make umbra thread-shared for improved scalability

_From [bruen...@google.com](https://code.google.com/u/109494838902877177630/) on May 16, 2012 13:01:27_

want per-thread "unicache" b/c different threads use same code to access
different data

perhaps a combination w/ adaptive scheme to figure out where per-thread
uni-cache is needed. we can start with thread-shared uni-cache. later if
one static instruction keeps accessing different unit, we will allocate a
TLS slot for it for thread-private uni-cache.

on linux can have tons of TLS slots.
but harder on windows: how find enough slots?
stealing a register is one solution.
it's even possible it's faster than using TLS b/c that goes through slow decoder
and expands instruction size.

_Original issue: http://code.google.com/p/dynamorio/issues/detail?id=775_