Use compact valuation mark columns

[tradingexpert]

Summary

• return C++ replay valuation marks as compact int64 byte columns instead of one MarkRow object per mark
• add CompiledValuationMarks and teach equity-curve construction to consume compact marks directly
• keep ReplayResult.equity_curve public output unchanged while avoiding intermediate ValuationMark object churn
• extend benchmarks with CppMatchingEngine batch+marks so the audited replay cost is easier to explain

Benchmarks

Direct engine:

• MatchingEngine scalar: 1,774,050 events/s
• CppMatchingEngine per-event: 1,844,263 events/s
• CppMatchingEngine batch: 16,980,531 events/s
• CppMatchingEngine batch+marks: 18,012,722 events/s
• batch+marks speedup vs Python: 10.15x

Full audited replay:

• Replay + Python engine: 374,845 events/s
• Replay + default engine: 561,666 events/s
• default engine speedup vs Python: 1.50x

Tests

• python -m pip install -e ".[dev]"
• python -m ruff check .
• python -m pytest
• mkdir -p build/cpp && c++ -std=c++17 -Wall -Wextra -pedantic -Icpp cpp/tests/matching_engine_core_test.cpp -o build/cpp/matching_engine_core_test && build/cpp/matching_engine_core_test
• python -m benchmarks.engine_throughput
• python -m benchmarks.replay_throughput

Coverage: 98.27% total.