Mark fake home directories that failed to be created, so we won't
try to remove them and have cleanup just use p_rmdir.

Without this change, any failed assertion in the second (or a later) test
inside a test suite has a chance of double deleting memory, resulting in
a heap corruption. See #1096 for details.

This leaves alone the test cases where we "just" use cl_git_sandbox_init()
and cl_git_sandbox_cleanup(). These methods already take good care to not
double delete a repository.

Fixes #1096

The new Win32 global path search was not working with the
environment variable tests.  But when I fixed the test, the new
codes use of getenv() was causing more failures (presumably because
of caching on Windows ???).  This fixes the global file lookup to
always go directly to the Win32 API in a predictable way.

Not found means not found, and the other way around.

gcc 4.7.0 apparently doesn't see that we won't call setenv with NULL as
second argument.

The error codes from failed lookups of system and global files
on Windows were not consistent with the codes returned on other
platforms.  This makes the error detection patterns match and
adds a unit test for the various errors.

This fixes two bugs:

* Issue #728 where git_status_file was not working for files
  that contain spaces.  This was caused by reusing the "fnmatch"
  parsing code from ignore and attribute files to interpret the
  "pathspec" that constrained the files to apply the status to.
  In that code, unescaped whitespace was considered terminal to
  the pattern, so a file with internal whitespace was excluded
  from the matched files.  The fix was to add a mode to that code
  that allows spaces and tabs inside patterns.  This mode only
  comes into play when parsing in-memory strings.

* The other issue was undetected, but it was in the recently
  added code to reload gitattributes / gitignores when they were
  changed on disk.  That code was not clearing out the old values
  from the cached file content before reparsing which meant that
  newly added patterns would be read in, but deleted patterns
  would not be removed.  The fix was to clear the vector of
  patterns in a cached file before reparsing the file.

The function to convert UTF-16 to UTF-8 was only allocating a
buffer of wcslen(utf16str) bytes for the UTF-8 string, but that
is not sufficient if you have multibyte characters, and so when
those occured, the conversion was failing.  This updates the
conversion functions to use the Win APIs to calculate the correct
buffer lengths.

Also fixes a comparison in the unit tests that would fail if
you did not have a particular environment variable set.

On Windows, we are having problems with home directories
that have non-ascii characters in them.  This rewrites the
relevant code to fetch environment variables as UTF-16 and
then explicitly map then into UTF-8 for our internal usage.