perf(bench): Wave 0 full-pipeline throughput on one A6000 (binding resource per stage)

[davidamacey]

Refs #45, #210, #230. Also opens #239, #240, #241.

Extends the v0.6.0 baseline from the 4-client COCO run to the whole pipeline on one RTX A6000 (GPU 2, idle, FP16 PE engine, v0.5.0 images, isolated project, loopback only), 3 repetitions, medians in docs/PERFORMANCE.md ("0.6.0 Wave 0: full-pipeline throughput"), raw JSON in docs/benchmarks/v060_wave0_*.json.

Headlines

• COCO frames, 2,000 images, 32 per request: 22.7 images/s upload, 22.8 batch at 16 clients, GPU 94 % busy (97 % of the 23.5 ceiling); 18.7 at the baseline's 4 clients. The GPU is power-capped (300 W, SM 1.42 to 1.50 GHz). Triton instance counts, queue delay and preferred_batch_size change nothing (within 3 %).
• 500 pinned public photos, 12 to 24 MP (new fetch_hires_set.py, commons_hires_500.json): host-bound, 1.0 to 1.3 API CPU-s per image (0.18 on COCO), 3 to 11 images/s, GPU 5 to 15 % busy; 98.5 % of API CPU is on the event-loop threads; throughput falls past 32 clients (perf(ingest): on 12-24 MP photos the API binds, not the GPU (1.2 CPU-s per image, 98 percent on the event-loop thread, throughput falls past 32 clients) #241).
• Faces cost 44 to 54 ms of API CPU per COCO image; OCR is capped at 21 images/s by one ocr_pipeline instance (42 with 2, 80 with 4, perf(triton): ocr_pipeline runs one instance, so OCR is capped at about 21 images/s (2 instances 42, 4 instances 80) #240).
• OpenSearch is far from binding: 1,011 documents/s from one writer (137 images/s equivalent), 6,987 from sixteen; ef_construction 128 cuts index CPU per document by 38 %.
• Not measured: region stage (gated segmenter weights), per-GPU scaling (the other GPUs host other projects), cluster pass.
• Defect found: an oversize upload answers 500, not 413, and the 50 MB cap counts the whole request (fix(api): oversize upload returns 500 instead of 413, and the 50 MB cap counts the whole multipart request #239).

Code: wave0_suite.py, wave0_probes.py, wave0_lib.py, wave0_report.py, os_write_ceiling.py, ingest_micro_profile.py; baseline_suite.py gains --skip-cluster, --skip-settle and a pluggable gauges class. Tests under tests/bench and tests/datasets.

Checks: pre-commit run --from-ref origin/main --to-ref HEAD exits 0, scripts/docs/check_docs_vs_code.py clean, pytest tests/bench tests/datasets passes.