Profile low level (OS) memory allocation

[pjenvey]

Under load testing scenarios for #76 we're seeing a significantly higher RSS size for autopush/endpoint processes vs the amount of memory allocated by PyPy.

PyPy's total allocated memory size has been confirmed via introspecting data structures in gdb and also via initial versions of PyPy's new mem stats API per [Q4 OKR 2.1](mozilla-services/push-okr#52

The only other significant allocations seen under massif profiling are from OpenSSL. It should however have reasonable memory usage after fixing autopush#1038.

We need extra memory statistics about the OS level malloc, because we don't seem to be 100% responsible for our memory leak.

PyPy GC's may cause some fragmentation over time but we're seeing way too much leakage/memory usage.

[pjenvey]

A brief report of profiling of glibc malloc via its malloc_stats(3): https://drive.google.com/open?id=1-DN_6nxm6RRyxmMmGw0VAyGA-UFLl8jjuRgpeKEadvU

Next steps:

• Test autopush/endpoint w/ low MALLOC_ARENA_MAX values
• Support jemalloc (optionally)