Change object marking routines to minimize stack usage.

Change object marking routines to minimize stack usage.
This change moves a few cold paths from mark_object to NO_INLINE
functions and adjusts symbol marking loop.  According to GCC 4.8.2
-Wstack-usage, this reduces mark_object's stack usage from 80 to
48 bytes on a 64-bit system.  For a long byte-force-recompile runs,
stack usage at the mark phase is reduced up to 28%.  Surprisingly,
it also gains up to 3% in speed (with default '-O2 -g3' flags).
* alloc.c (mark_compiled, mark_localized_symbol): New functions,
refactored out from ...
(mark_object): ... adjusted user.  Also mark symbols in a tight
inner loop.
(mark_face_cache): Add NO_INLINE.