#285 - Add direct polling latency benchmark and BlueFlame posting

[cliffburdick]

Reduce the latency on TX direct poll mode by writing directly to BF registers

[greptile-apps]

Greptile Summary

The PR adds a direct-polling latency benchmark and reduces direct ibverbs TX overhead through full BlueFlame posting, CPU-inline WQEs, batched completion reclamation, and lock-free queue ownership.

• Adds the latency benchmark executable, configuration template, CSV reporting, PTP clock calibration, and usage documentation.
• Adds boundary-safe staging for wrapped inline WQEs before full BlueFlame copies.
• Updates direct TX completion signaling, draining, queue ownership, and diagnostics.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
src/engines/ibverbs/daqiri_ibverbs_engine.cpp	Implements direct-queue ownership, inline and full BlueFlame posting, cyclic-SQ staging, batched completion reclamation, and explicit drain signaling; the previously reported SQ-boundary read is addressed.
src/engines/ibverbs/daqiri_ibverbs_engine.h	Adds the queue state and counters required for lock-free ownership, inline WQEs, completion cadence, and drain tracking.
examples/raw_latency_bench.cpp	Adds a direct-polling TX/RX latency benchmark with hardware RX timestamps, optional PTP cross-clock calibration, summaries, and CSV output.
examples/daqiri_bench_raw_latency_ibverbs.yaml	Provides the host-customizable direct ibverbs TX/RX configuration for the new latency benchmark.
examples/CMakeLists.txt	Builds, links, installs, and stages the new latency benchmark and configuration.
docs/benchmarks/raw_benchmarking.md	Documents benchmark setup, timing boundaries, PTP calibration, output metrics, and loopback operation.

_{Reviews (5): Last reviewed commit: "#285 - Clarify latency measurement timel..." | Re-trigger Greptile}