feat: add AIPerf multimodal benchmark integration

[ActivePeter]

Summary

• expose bounded target-side phase, chunk, runtime, and environment facts
• add TeleFuser video and LingBot stream benchmark profiles plus an SGLang baseline contract
• clone and use AIPerf only at <TeleFuser>/benchmarks/aiperf, with no checkout-path override
• keep warmup, aggregation, semantic mapping, artifacts, resource telemetry, GreptimeDB, and UI ownership in AIPerf
• document the user workflow and metric protocol in English and Chinese

Validation

• changed-area unit tests: 4 passed
• broader unit run: 754 passed and 7 skipped; 6 existing pipeline-parallel tests require CUDA on the local runner
• integration run: 44 passed, 1 skipped, 2 GPU tests deselected
• pre-commit passed
• strict MkDocs build passed
• live 2xH100 LingBot WebRTC quick profile succeeded with 6 steady chunks, 4.3565 target compute FPS, and 16.5361 client stream FPS
• active resource telemetry merged CPU, memory, GPU, GPU memory, Ethernet, and RDMA curves into the same GreptimeDB run with zero dropped samples

Metric boundary

LingBot compute time spans actor submission through raw-frame return. Native WebRTC encode time remains unavailable. Child-actor allocator peaks are omitted instead of reporting incomplete service-process values; AIPerf process-tree GPU-memory telemetry provides the runtime curve.

Companion AIPerf PR: ActivePeter/aiperf#1

[ActivePeter]

Final validation for 4affe5a completed:

• pre-commit run --all-files: passed
• CI-equivalent CPU unit selection: 675 passed, 2 skipped
• MkDocs strict build, shell syntax, JSON/JSONL, and YAML validation: passed
• fresh GitHub clone + documented scripts/setup_aiperf_repo.sh: passed; it installed teleai@e977ffbb under benchmarks/aiperf and created artifacts/
• GreptimeDB + History API/bundled Vue UI smoke test: passed; a real TeleFuser artifact imported 64 metric points and the repeat import was skipped idempotently

The Wan I2V example now defers the GPU-dependent pipeline import, fixing the two torchvision::nms failures seen on the prior CI SHA. The latest fork workflows are awaiting maintainer approval: Lint and Tests.

[lzx1413]

LGTM