base: shrink the code FlatHashMapV2 generates

[LalitMaganti]

Summary

Cuts the code FlatHashMapV2 generates per call site to under what
FlatHashMapV1 generates, without slowing down inserts.

Why

FlatHashMapV2 force-inlined its whole probe loop into every Find() and
Insert(): ~750 bytes of code per insert+find call site vs ~150 for
FlatHashMapV1. Moving code over to it, which is otherwise faster, would
have grown binaries by hundreds of KB.

Changes

• Insert() inlines only constructing the key and value. Finding or
  claiming the slot is out of line in FindOrPrepareInsert(), compiled
  once per key type. Find() is out of line too.
• The state and the operations which don't depend on the key and value
  types (allocating, growing, probing for a free slot) move to an
  untemplated FlatHashMapV2Base, shared by all maps, defined in the
  new src/base/flat_hash_map.cc.
• Rehashing moves entries straight into their new slots, rather than
  through Insert().
• Lookups work as in absl: Find(), Erase(), Insert() and
  operator[] take a Key, which the caller converts to, and only a
  transparent Hasher enables other (heterogeneous) key types.
  MurmurHash is now transparent only for string-like keys: it also was
  for integers, hashing an int looked up in a map keyed by uint32_t
  as an int, so e.g. a negative one could miss (in FlatHashMapV1 too).

Benchmarks

Code size for 16 map types (x64, -O3):

	1 site per type	8 sites per type	Per insert+find site
FlatHashMapV1	26.9 KB	44.0 KB	~150 B
FlatHashMapV2 before	37.3 KB	121.1 KB	~750 B
FlatHashMapV2 after	~27 KB	~35 KB	~80 B

Speed, flat_hash_map_benchmark on Linux x64: inserts are unchanged.
Lookups through the out of line Find() are 4-14% slower than before,
still ~1.7x faster than FlatHashMapV1.

Testing

FlatHashMap* and MurmurHash* unit tests, with one test changed to
construct its Key (an int is no longer a heterogeneous key for it) and
one added for MurmurHash's transparency. A trace processor build moved
onto this (in the follow-up PR) gives the same query results on three
traces.

[github-actions]

🎨 Perfetto UI Builds & Tests

• 04de724 (2026-10-08 21:51 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-37848479666-1-ui/ui/index.html
• 5c35237 (2026-10-08 22:42 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-37854266906-1-ui/ui/index.html
• d220c98 (2026-10-08 23:08 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-37856626113-1-ui/ui/index.html
• a44d1a5 (2026-10-08 23:19 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-37857949980-1-ui/ui/index.html
• cc7e1a0 (2026-10-09 17:03 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-37962517159-1-ui/ui/index.html

[safayat-google]

Looks Great!

Just few nits.