load-testing: add MLflow tracing structure, validation + bottleneck examples

[RamVegiraju]

Summary

• Documents that the load-testing skill assumes the app is served by the MLflow AgentServer (the /invocations SSE contract the scripts hit, and the source of the per-request traces).
• Adds Step 6 — Validate & Diagnose with MLflow Tracing: reliability from trace state, latency over successful traces, and bottleneck diagnosis via the span waterfall (CHAT_MODEL vs TOOL split, per-tool p95). Locust measures the system from the outside, MLflow traces from the inside — together they turn a QPS number into an explanation.
• Adds two starter examples: examples/locustfile.py (streaming, TTFT, M2M OAuth, ramp shape) and examples/validate_with_mlflow.py (reliability + latency + span breakdown).
• Adds a Resources section (Locust, MLflow tracing/search-traces/entities API, AgentServer, Databricks load-test docs) and an MLflow-tracing trigger to description.
• Fixes a tracking gap: adds load-testing to the .gitignore source-skill allowlist so its example files are versioned and sync-skills.py can reproduce the template copies from a clean clone.

All changes are additive (existing wording untouched aside from the extended description) and synced to all templates via .scripts/sync-skills.py.

Test plan

• Both example scripts run end-to-end against a real experiment; validate_with_mlflow.py correctly surfaced the bottleneck tool (get_stock_price p95 ≈ 897ms vs get_weather ≈ 200ms) and the model-vs-tool split.
• APIs validated against installed OSS source + live data: Locust events.request.fire / LoadTestShape.tick; MLflow search_traces(locations=…), trace.data.spans, span_type/start_time_ns/end_time_ns.
• examples/*.py compile in source and all synced template copies.
• Doc claims validated against mlflow.org (span types CHAT_MODEL/TOOL/CHAIN, leaf-span caveat).