docs(rocke): single kernel-authoring DoD, agent rules, and check runner (#10380)

## Summary

Establishes a single, consolidated engineering process and Definition of
Done for rocke kernel authoring, plus one auto-discovering local check
runner and a canonical agent-rules entry point. Docs + tooling only — no
kernel/emitter or dispatcher changes.

- **`KERNEL_AUTHORING.md`** — the one process + DoD doc: Core DoD (every
change), per-change-type checklists (new kernel / new knob / new feature
/ perf), a local test matrix, and an attention (SDPA/MHA)
library-process appendix. Generic core so sibling kernel teams can adopt
it and add their own appendix.
- **`AGENTS.md`** (with `CLAUDE.md`, `.clinerules`,
`.cursor/rules/ai-rules.mdc` as symlinks) — the canonical agent-rules
index: compliance, the byte-identity invariant, layout, hard rules, and
optimization-doc routing. One source of truth so tool-specific rule
files never drift.
- **`README.md`** — top-level rocke landing page and "start here"
routing table.
- **`tools/run_checks.py`** — one command that auto-discovers every
`test_*.py` and `<family>_emit.{py,c}` parity pair (no manual
registration). Runs guard → byte-identity gate (both LLVM flavors) →
platform + library parity → pytest → on-GPU numeric. Scope with `--steps
<subset>` and `--op <operator>` (e.g. `--steps numeric --op fmha_bwd`).
- **`platform/dsl_docs/instances/attention.md`** — refreshed to the
post-#8977 `platform/` + `library/` layout (kernel/builder/dispatch
paths, current family list).
- **`.gitignore`** — ignore local-only files and build artifacts under
the rocke tree.

## Risk

Low. Additive docs + standalone tooling under
`dnn-providers/hip-kernel-provider/rocke/`; nothing wired into the
build, CI, or dispatch. The only modified existing file is
`attention.md` (doc-only path/layout refresh).

## Testing

- `python tools/run_checks.py --list` — auto-discovers the library
parity pairs + both pytest suites; GPU detection works.
- All internal doc links validated to resolve within the repo tree; no
dangling references to the consolidated-away docs.
- All Python parses; no scratch/artifact/local-only files included.

## Note

rocke CI is owned separately by the CI team and is intentionally out of
scope here; `run_checks.py` is the developer-run local gate in the
meantime.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude <noreply@anthropic.com>