stats: use flat_hash_map and make thread-local stat maps use key storage from stat

[jmarantz]

Description: A significant amount of storage is wasted in the maps in ThreadLocalStore with copies of the strings. This changes the maps to use const char* keys, which are held in every form of stat. This PR also switches the maps in thread-local-store to use flat_hash_map, which is particularly effective when the keys are just pointers, #4711 and #4715 are much clearer wins when submitted together, at the cost of complexity of review. This is one step toward resolving #4196.

It reduces memory from 73k/cluster to 56k/cluster in the Istio test referenced in #4196. In the memory-tests in thread_local_store_test.cc, it reduces the no-tls test from 31.5M to 21.2M, and the TLS memory test from 40.4M to 24.5M.

This change also makes things much faster:

current:

--------------------------------------------------------------
Benchmark                       Time           CPU Iterations
--------------------------------------------------------------
BM_StatsNoTls_mean       16249144 ns   16248983 ns         37
BM_StatsNoTls_median     16176011 ns   16175443 ns         37
BM_StatsNoTls_stddev       484905 ns     484909 ns         37
BM_StatsWithTls_mean     20958468 ns   20958244 ns         30
BM_StatsWithTls_median   20899829 ns   20899217 ns         30
BM_StatsWithTls_stddev     576702 ns     576784 ns         30

new (this PR):

--------------------------------------------------------------
Benchmark                       Time           CPU Iterations
--------------------------------------------------------------
BM_StatsNoTls_mean       13779759 ns   13779854 ns         39
BM_StatsNoTls_median     13630165 ns   13630325 ns         39
BM_StatsNoTls_stddev       576447 ns     576471 ns         39
BM_StatsWithTls_mean     16336826 ns   16336967 ns         35
BM_StatsWithTls_median   16240568 ns   16240563 ns         35
BM_StatsWithTls_stddev     544943 ns     544942 ns         35

A significant follow-up to this would be to tackle the redundant storage for tagExtractedName and TagVector. Another significant follow-up is to use symbolized strings (see #4696 )

Risk Level: low
Testing: //test/...
Docs Changes: n/a
Release Notes:n/a

[mattklein123]

Very cool. LGTM, small comment.