Add benchmark suite for engine and replay throughput

[tradingexpert]

What changed

• add shared mixed MBO benchmark workloads
• add direct engine throughput benchmark for Python scalar, C++ scalar, and C++ batch paths
• add full audited replay throughput benchmark for Python reference and default replay paths
• document how to interpret direct engine throughput separately from full replay throughput
• clarify that ordinary audited Replay currently remains event-by-event, while direct C++ batch ingestion is the fast path for callers that own the event loop

Why

After #46, the old scalar-only benchmark story was stale. This PR makes the performance boundary visible: the direct C++ batch path is genuinely faster, while the current audited Replay path still pays for per-event gateway and valuation behavior. It also records the next intended performance step: boundary-batched replay that lets C++ run independently until Python must observe a fill, instruction boundary, or configured inspection point.

Local benchmark results

Using --cycles 20000 --repeats 3 --warmups 1 on this machine:

Direct execution-engine throughput
MatchingEngine scalar          120,000 events    0.0657 s     1,826,506 events/s
CppMatchingEngine scalar       120,000 events    0.0627 s     1,914,635 events/s
CppMatchingEngine batch        120,000 events    0.0069 s    17,517,716 events/s
scalar C++ speedup vs Python     1.05x
batch C++ speedup vs Python      9.59x

Full audited replay throughput
Replay + Python engine         120,000 events    0.2865 s       418,829 events/s
Replay + default engine        120,000 events    0.2855 s       420,242 events/s
default engine speedup vs Python     1.00x

A private research feed-path smoke comparison on the same broad mixed workload shape measured about 1.7M events/s; it is not apples-to-apples with public audited Replay, but it confirms where the public replay bridge still needs work.

Validation

• python -m ruff check .
• python -m pytest
• python -m benchmarks.engine_throughput --cycles 20000 --repeats 3 --warmups 1
• python -m benchmarks.replay_throughput --cycles 20000 --repeats 3 --warmups 1