[RWKV] Align community integration contracts

[Cai-z-us]

What does this PR do?

Aligns the RWKV-7 integration with the agreed community contracts while keeping the canonical RWKV math, tensor names, head_size=64, and FlashRWKV2 provider boundary unchanged.

• move linear-attention cache batch operations into the shared cache layer and invalidate stream-bound metadata after state reordering
• enable standard Transformers gradient checkpointing for stateless and stateful training
• centralize the equal-length, all-ones attention-mask contract and preserve standard generation input preparation
• document compile, export, TP, batching, and inference-layout support boundaries
• make the in-place prepare_for_inference() lifecycle explicit and test the BF16 training round trip
• enforce and test the unconditional fixed-commit tokenizers-rwkv dependency and RwkvTrie save/reload contract

Qwen2Rwkv remains intentionally out of scope.

Coordination

No repository issue is mapped to this maintenance batch; it was requested and approved directly by the repository owner. Duplicate-work checks found no overlapping open PR or issue in this repository. Upstream already has a separate portable RWKV-7 proposal in huggingface#47780, so this PR is deliberately limited to the rwkv-rs integration branch and does not create a competing upstream PR.

Validation

• local targeted suite: 66 passed, 1 skipped (the skipped case requires the official vocabulary path)
• rwkv-sha-pro6000x8 with the official vocabulary: 67 passed, 9 subtests passed
• make typing
• no-cache modeling structure checks
• RWKV checks with the historical contribution cutoff disabled
• repository consistency checks
• wheel metadata check for the exact Git tokenizer commit and FlashRWKV2-only rwkv extra
• make style
• git diff --check

Before submitting

• Read the contributor and PR checks
• Updated the documentation for the public support contract
• Added the necessary regression tests

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33049539414
Result: failure | Grafana metrics are not available yet.

Code quality check failed: test jobs were skipped. Fix the code quality issues and push again to run tests.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: rwkv

[Cai-z-us]

Post-merge CI note: Build PR Documentation completed successfully. The three red PR CI preparation jobs never started a runner; each GitHub annotation reports Required runner group 'aws-m8i-l-cache' not found. This is an unavailable private runner group in the fork's inherited reusable workflow, not a code/test failure. The PR was merged after the documented local checks and the full rwkv-sha-pro6000x8 suite (67 passed, 9 subtests passed).