Also fixes issue with unicode callbacks never being set really.

Patch from Jonathan Kew.

Kinda hand-wavy right now.  Not tested.

Fixes lack of head_table initialization in create_for_tables() also.

We plan to use the first bit to be always on.

Before, the mask in the buffer was inverted.  That is, a 0 bit meant
feature should be applied and 1 meant not applied, whereas in the
lookups, the logic was positive.

Now both are in sync.  When calling hb_buffer_add_glyph() manually,
the mask should be 1 instead of 0.