test: TransTableL op-stats unit tests (#379 item 3)

[wopdevries]

Adds unit test coverage for TransTableL's op-stats counters (num_adds_, num_overwrites_, num_harvests_), instrumented in #365 but never directly tested — item 3 from #379.

Tests added (library/tests/trans_table/trans_table_l_op_stats_test.cpp):

InsertIncrementsAddsOnly — a single new entry increments adds only.
FullBlockReplacementIncrementsOverwrites — filling a block to BlocksPerEntry (125) capacity and adding one more distinct entry increments both adds and overwrites.
ResetOpStatsClearsCounters — reset_op_stats() zeroes all three counters after activity.

Implementation note: the public API has no way to fill a WinBlock to capacity, so the full-block test seeds WinBlock/WinMatch state directly via a friend class TransTableLOpStatsTest; declaration added to trans_table_l.hpp, rather than reverse-engineering add()'s internal win-rank/aggregation encoding to manufacture 125 organic collisions. Seeded entries use mask_index_ = 0 / top_set1_..4_ = 0, a combination the real encoding never produces once any suit has a nonzero win rank — so the real add() call in the test is guaranteed to take the "not found → new insert" path rather than accidentally matching a seeded entry.

Verification: compiled and ran against the real sources outside Bazel first, then confirmed each test actually catches its corresponding regression by deliberately disabling num_adds_++/num_overwrites_++ and separately reset_op_stats()'s body — each break fails exactly the test(s) it should and no others. Also compiled alongside the other three files sharing this Bazel target (trans_table_base_test.cpp, trans_table_s_test.cpp, trans_table_p_test.cpp) to rule out link-time collisions. bazelisk test //library/tests/trans_table:trans_table passes on the real toolchain.

Refs #379 (item 3 of 4 — items 1/2 covered by #397/#402; item 4, the analyse_later_board reset/report scoping question, is a separate design discussion still in progress).

[zzcgumn]

LGTM