[tool] feat: add profile skill

[NancyFyong]

What does this PR do?

Adds a new agent skill profile under .agents/skills/profile/SKILL.md.

It is a router (same style as add-pipeline / run-cpu-tests / self-review):
it points at the authoritative docs/perf/profiler.md and fills only the
cross-cutting decisions that guide leaves implicit — which tool answers which
question (nsys / torch.profiler / torch_memory / MFU / RL-Insight), why to
profile one lightweight step instead of a full run, which process owns each phase
(actor / rollout server / reward-model server / ref), the key=value vs
+key=value override rule, the main_diffusion_v1 nsys caveat, and not
hand-rolling profilers in adapter code. No code, config, or docs pages change.

Checklist Before Starting

• Searched open/merged PRs: this does not duplicate [cfg] feat: support NPU profiling #507 ([cfg] feat: support NPU profiling, which adds NPU profiling config code) or [doc] feat: add performance tuning guide entry point #487 ([doc] feat: add performance tuning guide entry point, a docs page). No PR adds a
  profiling agent skill under .agents/skills/.
• Title follows [{modules}] {type}: {description} ([tool] feat: ...),
  matching the precedent set by [tool] feat: add self-review skill ([tool] feat: add self-review skill #518).

Test

No code paths change, so there are no unit tests. Verified locally:

• All 12 pre-commit hooks pass on the changed file (pre-commit run --files .agents/skills/profile/SKILL.md), including Generate and verify _generated_*.yaml (no config touched), Check docs Last updated info, and
  Check license — SKILL.md is correctly excluded from both, like the
  existing skills.
• PR-title validator passes: PR_TITLE="[tool] feat: add profile skill" python tests/special_sanity/check_pr_title.py → valid, modules ['tool'], type
  feat.
• All three docs the skill routes to exist: docs/perf/profiler.md,
  docs/perf/diffusion_mfu.md, docs/start/rl_insight.md.

API and Usage Example

Discovered by the agent harness the same way as the other skills (one
SKILL.md per directory with name + description front matter). It triggers
when a performance/profiling question comes up, e.g.:

> profile the actor update of the SD3.5 OCR run and tell me where the time goes
# -> loads .agents/skills/profile/SKILL.md, which routes to docs/perf/profiler.md
#    recipe 1 (torch) or 4a (nsys), pins a one-step lightweight footprint, and
#    enables actor_rollout_ref.actor.profiler.

Design & Code Changes

• New file: .agents/skills/profile/SKILL.md (front matter + a concise routing
  body). Nothing else changes.
• Deliberately a router, not a copy: docs/perf/profiler.md stays the single
  source of truth for the config surface and the six recipes; the skill only
  adds tool-selection and the recurring traps.

Checklist Before Submitting

• Read the AGENTS.md.
• Follows the [{modules}] {type}: {description} title format.
• Pre-commit hooks pass.
• Human line-by-line review of the changed file.

AI assistance was used for this change.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.