perf(decoder): add bounded columnar decoding

[muthmann]

Summary

• decode EVT3 data directly into columnar storage and stable NumPy arrays
• add bounded-memory RAW and HDF5 file readers, including trigger-aware Python batches
• stream CLI CSV and binary output without materializing the full recording
• optimize sparse vector masks, byte conversion, capacity planning, GIL use, and release profiles
• add Criterion coverage, a PGO build workflow, measured optimization results, and updated README/feature documentation
• include the NumPy-native Events and Augur ingress foundation required by the optimized Python workflow
• retain and strengthen the C++ reference comparison with a reproducible like-for-like CSV benchmark

User impact

Existing decode_file, decode_file_with_triggers, and decode_bytes calls remain compatible and use the optimized core automatically. Large-file users can opt into decode_file_batches; users who need external trigger edges can use decode_file_batches_with_triggers, which yields (Events, TriggerEvents) per bounded batch.

Measured on test_data/laser.raw:

• Python full-file path: 1.82x faster and 59.1% lower maximum RSS than the previous implementation
• bounded Python path: 92.1% lower maximum RSS than the previous full-memory path
• streaming CLI: 1.67x faster and 93.2% lower maximum RSS than the previous materialized CLI
• sparse vector-mask benchmark: about 1.89x throughput
• like-for-like full CSV: Rust 7.414 s versus optimized C++ 12.028 s, or 1.62x faster
• instrumented CSV run: Rust 63.3 MB RSS versus C++ 28.3 MB; C++ remains more memory-efficient in this comparison

The Rust and C++ CSV outputs were byte-identical on an 8-MiB input prefix. The benchmark runner now verifies this and alternates measured run order.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• HDF5 Clippy and full HDF5 test suite
• release Python extension build
• Python test suite: 21 passed
• reproducible Rust/C++ comparison mode executed successfully
• byte-identical Rust/C++ CSV prefix with SHA-256 398d63a52eeb7291caa1346037f674ef0b6209772d3c512e2aa6c70b9aed12f4
• exact aggregate checksum comparison for the real-file Python result
• PGO build and output comparison
• local Markdown-link audit

Real ECF fixture checks were skipped because the external ECF plugin is not installed. Synthetic HDF5 and HDF5 batch tests passed.