odb_pack: make sure to search all backends

last_found is the last packfile a wanted object was found in. Since
last_found is shared among all searching threads, it might changes while
we're searching. As suggested by @arrbee, put a copy on the stack to fix
the race condition.

refs: prevent locked refs from being enumerated

Fix #936

Defining the BOM as a string makes the array include the
NUL-terminator, which means that the memcpy is going to check for that
as well and thus never match for a nonempty file.

Define the array as three chars, which makes the size correct.

I don't think clone.c needs in #include dirent.h and it is not
portable, so let's just get rid of it.

Fix MSVC compilation warnings

pack: iterate objects in offset order

Compute the ordering on demand and persist until the index is freed.

The winhttp branch was based on a version before these existed, so the
build broke on Windows.

http: use WinHTTP on Windows

Wondows has its own HTTP library. Use that one when possible instead of
our own.

As we don't depend on them anymore, remove the http-parser library from
the Windows build, as well as the search for OpenSSL.

Fix error text s/buffer too long/buffer too short/

Fix problems in diff iterator record chaining

The last change tweaked the way we use the hunk_curr pointer
during iteration, but failed to reset the value back to NULL
when switching files.

refspec: No remote tracking ref from a fetchspec-less remote

There is a bug in building the linked list of line records in the
diff iterator and also an off by one element error in the hunk
counts.  This fixes both of these, adds some test data with more
complex sets of hunk and line diffs to exercise this code better.

Add git_repository_hashfile to hash with filters

odb_pack: try lookup before refreshing packs

This reduces the rate of syscalls for the common case of sequences of
object reads from the same pack.

Best of 5 timings for libgit2_clar before this patch:
real    0m5.375s
user    0m0.392s
sys     0m3.564s

After applying this patch:
real    0m5.285s
user    0m0.356s
sys     0m3.544s

0.6% improvement in system time.
9.2% improvement in user time.
1.7% improvement in elapsed time.

Confirmed a 0.6% reduction in number of system calls with strace.

Expect greater improvement for graph-traversal with large packs.

Fixed some minor `git_repository_hashfile` issues:

- Fixed incorrect doc (saying that repo could be NULL)
- Added checking of object type value to acceptable ones
- Added more tests for various parameter permutations

The existing `git_odb_hashfile` does not apply text filtering
rules because it doesn't have a repository context to evaluate
the correct rules to apply.  This adds a new hashfile function
that will apply repository-specific filters (based on config,
attributes, and filename) before calculating the hash.

git_mergebase: Constness-Fix for consistency

In the process of adding tests for the max file size threshold
(which treats files over a certain size as binary) there seem to
be a number of problems in the new code with detecting binaries.
This should fix those up, as well as add a test for the file
size threshold stuff.

Also, this un-deprecates `GIT_DIFF_LINE_ADD_EOFNL`, since I
finally found a legitimate situation where it would be returned.

cache: fix race condition

Example: a cached node is owned only by the cache (refcount == 1).
Thread A holds the lock and determines that the entry which should get
cached equals the node (git_oid_cmp(&node->oid, &entry->oid) == 0).
It frees the given entry to instead return the cached node to the user
(entry = node). Now, before Thread A happens to increment the refcount
of the node *outside* the cache lock, Thread B tries to store another
entry and hits the slot of the node before, decrements its refcount and
frees it *before* Thread A gets a chance to increment for the user.

	git_cached_obj_incref(entry);

	git_mutex_lock(&cache->lock);
	{
		git_cached_obj *node = cache->nodes[hash & cache->size_mask];

		if (node == NULL) {
			cache->nodes[hash & cache->size_mask] = entry;
		} else if (git_oid_cmp(&node->oid, &entry->oid) == 0) {
			git_cached_obj_decref(entry, cache->free_obj);
			entry = node;
		} else {
			git_cached_obj_decref(node, cache->free_obj);

// Thread B is here

			cache->nodes[hash & cache->size_mask] = entry;
		}
	}
	git_mutex_unlock(&cache->lock);

// Thread A is here

	/* increase the refcount again, because we are
	 * returning it to the user */
	git_cached_obj_incref(entry);

Fix logical error in git_index_set_caps

git reference peel

Often `git_odb_read_header` will "fail" and have to read the
entire object into memory instead of just the header.  When this
happens, the object is loaded and then disposed of immediately,
which makes it difficult to efficiently use the header information
to decide if the object should be loaded (since attempting to do
so will often result in loading the object twice).

This commit takes the existing code and reorganizes it to have
two new functions:

- `git_odb__read_header_or_object` which acts just like the old
  read header function except that it returns the object, too, if
  it was forced to load the whole thing.  It then becomes the
  callers responsibility to free the `git_odb_object`.
- `git_object__from_odb_object` which was extracted from the old
  `git_object_lookup` and creates a subclass of `git_object` from
  an existing `git_odb_object` (separating the ODB lookup from the
  `git_object` creation).  This allows you to use the first header
  reading function efficiently without instantiating the
  `git_odb_object` twice.

There is no net change to the behavior of any of the existing
functions, but this allows internal code to tap into the ODB
lookup and object creation to be more efficient.

This commit adds a max_size value in the public `git_diff_options`
structure so that the user can automatically flag blobs over a
certain size as binary regardless of other properties.

Also, and perhaps more importantly, this moves binary detection
to be as early as possible in the diff traversal inner loop and
makes sure that we stop loading objects as soon as we decide that
they are binary.

The `git_diff_iterator_num_files` API was problematic, since we
don't actually know the exact number of files to be iterated over
until we load those files into memory.  This replaces it with a
new `git_diff_iterator_progress` API that goes from 0 to 1, and
moves and renamed the old API for the internal places that can
tolerate a max value instead of an exact value.

When `git_submodule` became an opaque structure, I forgot to add
accessor functions for the fetchRecurseSubmodules config setting.
This fixes that.

Previously when diffing blobs, the diff code just ran with a NULL
repository object. Of course, that's not necessary and the test
for a NULL repo was confusing. This makes the blob diff run with
the repo that contains the blobs and clarifies the test that it
is possible to be diffing data where the path is unknown.