Apparently some AIX versions have 64bit pointers yet 32bit double
alignment.  Fix assertion.

to get the following recipe working:

<match>
  <test1 .../>
  <edit1 .../>
  <test2 .../>
  <edit2 .../>
</match>

as:

<match>
  <test1 .../>
  </edit1 .../>
</match>
<match>
  <test1 .../>
  <test2 .../>
  <edit2 .../>
</match>

This change reverts 9acc14c3
because it doesn't work as expected when building
with -fshort-enums which is default for older arms ABIs

Thanks for pointing this out, Thomas Klausner, Valery Ushakov, and Martin Husemann

After this change, the following works as expected:
$ FC_DEBUG=4 fc-match ":family=foo bar, sans-serif"
...
FcConfigSubstitute Pattern has 3 elts (size 16)
        family: "foo bar"(s) "sans-serif"(s)
...

Workaround to not failing even when the hash is unable to generate from fonts.
This change also contains to ignore the case if the hash isn't in either both
patterns.

based on hb-atomic-private.hh in harfbuzz

reflect correct path where is configured at the build time.

Regex is expensive to compare filenames. we already have the glob matching
and it works enough in this case.

Prior to this change, renaming FcConfigGlobMatch() to FcStrGlobMatch() and moving to fcstr.c

Add back FcHashGetSHA256DigestFromFile() and fall back to it
when font isn't SFNT-based font because FT_Load_Sfnt_Table
fails with FT_Err_Invalid_Face_Handle.