book: cache top-kDepthLevels Level* per side to cut publish_depth cost

[MustafaNazeer]

Summary

Measurement-driven follow-up to v1.1. With observability=true, the bench drops from 4.79 to 2.79 M evt/s on the desktop reference, a 142 ns/event tax. A per-piece bisection (toggle each side-effect, re-bench) attributes 90 ns of that to Book::publish_depth: an 8-step std::map iteration over the bid and ask price maps on every accepted event, plus a ~400 byte DepthSnapshot copy through the seqlock.

This PR makes Book carry a top-kDepthLevels Level* cache per side, updated by Book::add when a new Level is emplaced and by Book::remove_by_id / Book::erase_empty_level when a Level is erased. Book::publish_depth walks the cache (cache-warm objects the matching thread just touched) instead of the map. No wire format change; no thread-safety change; the existing depth seqlock still anchors reader correctness.

Measured paired runs on the desktop reference (same seed, same 1M events, same 100k warmup):

Mode	Pre-cache	Post-cache	Delta
observability=false (matching only)	4.79 M evt/s	4.60 M evt/s	-4%
observability=true (depth + trades + histogram)	2.79 M evt/s	3.39 M evt/s	+21%

Remaining observability overhead drops from 142 ns/event to 78 ns/event (about 45 percent reduction). p50 latency on the instrumented path drops from 288 to 230 ns. The observability=false regression is the cache maintenance cost on the Book::add / Book::remove_by_id hot path (linear scan plus shift, at most kDepthLevels operations) and remains comfortably above the 4.10 M evt/s CI baseline floor.

Cache correctness is exercised by seven new unit tests in tests/unit/test_book_depth_cache.cpp, including a 2000-step random-sequence test that audits the cache after every mutation via a new Book::audit_depth_cache_for_test() accessor (walks the maps and asserts cached pointers match the first kDepthLevels entries of each side). All 221 C++ tests pass on release build; existing property invariants P11 and P12, the byte-for-byte Differential test against the Python reference, and the SeqlockDepth concurrency test all continue to pass against the new publish_depth path.

Test plan

• Full C++ test suite green: ctest --test-dir build-release 221/221.
• Bench regression check vs bench/baseline.json: throughput -1.2%, p50 +9.4%, p99 +9.9%, p99.9 -41% (all within tolerance).
• Paired observability=on/off measurement captured in docs/perf/budget.md.
• CI green on both clang and gcc matrix jobs.
• Manual smoke test of the live demo after merge: Ladder, Tape, DepthChart, PerfPanel histogram continue to render real data.