Add compiled batch ingestion for C++ engine

[tradingexpert]

What changed

• add a public CompiledEventColumns helper that turns canonical MBOEvent rows into primitive columns
• add a compiled CppMatchingEngine.apply_events_batch(...) path over contiguous buffer inputs
• keep ordinary Replay(...) event-by-event so per-event valuation marks and equity curves stay unchanged
• document the distinction between replay semantics and direct compiled batch ingestion

Why

The private research engine gets its real C++ throughput from batch ingestion, while the public extraction only had one Python-to-C++ crossing per event. This ports the missing capability without weakening the inspectable replay contract.

Validation

• python -m ruff check .
• python -m pytest
• native C++ core test binary
• local targeted parity check against the private futures engine on the queue-fill scenario from PR Fix Databento trade normalization #45
• local throughput smoke test: public compiled batch path ~28.6M events/sec vs public scalar ~1.9M events/sec on this machine; private research batch path ~33.7M events/sec on the same broad workload shape