investigate and reduce redzone filling perf hit for pattern mode

_From [bruen...@google.com](https://code.google.com/u/109494838902877177630/) on April 11, 2012 16:03:47_

the actual memset of the redzone to the pattern needs to write 4x as much memory as a shadow set at default mapping.

1) need to investigate whether this explains some slowdowns in issue #750 , like bzip2 which was 2.9x in c2 but now it's 3.8x in c11

2) if it is an issue, come up with schemes to reduce the overhead

here are some ideas:
- could have sideline thread fill to-be-handed-out heap w/ pattern to
avoid filling at alloc time
- if we used our own allocator w/ per-size free lists we could avoid
re-filling pattern on re-use (xref issue #794 )

_Original issue: http://code.google.com/p/drmemory/issues/detail?id=859_