Fix some memory_order parameters in AtomicHashMap

Fix some memory_order parameters in AtomicHashMap

Summary: Herb Sutter noticed an error in the AHM find pseudo-code at
the C++ conference, and the error is in the real implementation too.
We should be using memory_order_acquire, because we need the load of
the key to happen before any subsequent loads of the value itself.
Our implementation is conservative for a memory_order_relaxed load
(we generate the same code as if it were memory_order_acquire---i.e. a
compiler barrier but no memory barriers), so this should not change the
generated code.  Correctness is a good idea, though (when we upgrade
to gcc 4.7's atomics it might not be as conservative).  Also, erase was
doing three loads when one is enough.  ... also fixes an exception safety
bug in insert.  (Phew.)

Test Plan: Ran the AHM unit tests, debug and opt.

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D485744