Rewrite review bot as a maintainer, not a counter

[lishunyang12]

Pivot the review bot from 'bot with comment caps' to 'maintainer with a checklist'. Based on vllm-omni's own docs/contributing/ (title prefixes, DCO, 5-level tests, RFC gate) plus the vllm-omni-skills catalog.

What changed

1. Three tiers, three real jobs (not one job at three depths):

   Label	Model	Job
   claude-quick	Haiku 4.5	Preview — summary + meta-check + file overview + depth recommendation. No inline bugs.
   claude-review	Sonnet 4.6	Full maintainer review
   claude-deep	Opus 4.7	Same scope as claude-review, stronger reasoning

2. Dropped hard caps. No '2-6 comments max'. A PR that warrants 15 gets 15; a PR that warrants 0 gets 0 and an empty approve.

3. Phase 0–3 maintainer flow (full review mode):

   • Phase 0 — PR meta audit (title prefix, DCO, description completeness, RFC gate)
   • Phase 1 — Category-specific checks keyed off the title prefix ([Model], [Quantization], [Kernel], [Hardware], etc.) — each has its own checklist drawn from docs/contributing/
   • Phase 2 — Code review (correctness / API / perf / scope creep)
   • Phase 3 — Test audit (mirrored paths, pytest markers, correct L-level)

4. No @claude trigger. Single mechanism = labels.

Rationale: a Haiku 'LGTM' isn't valuable — a Haiku preview is. Capping Sonnet/Opus at 6 comments makes the bot useless on big PRs. Real vllm-omni maintainers follow a category-specific checklist; the bot should too.