Thread-caching added to jemalloc

[Jason Evans]

I committed a largish set of jemalloc changes today, including the 
addition of thread-specific caching.  This has the potential to 
dramatically reduce the amount of locking necessary, but it can also 
increase memory usage.

I'm interested in feedback on how well the code works for your 
real-world multi-threaded applications.  If you do perform benchmarks, 
don't forget to define MALLOC_PRODUCTION in src/lib/libc/stdlib/malloc.c 
before building, and see the malloc(3) man page for tuning information. 
  In particular, note that you can disable thread-specific caching by 
setting MALLOC_OPTIONS=g in the environment, so it should be easy to 
perform comparative benchmarks.

Thanks,
Jason