feat(moore): support paged flash attention decode

[voltjia]

Summary

• Add a native Moore paged decode provider for FlashAttnWithKvcache.
• Register it as Moore implementation slot 8.
• Reuse the internal read-only paged-attention decode kernel without registering or calling the deprecated operator API.
• Cover FP16/BF16, head dimensions 64/128, MHA/GQA, tensor cache lengths, ALiBi, arbitrary outer strides, dynamic metadata, and external streams.

Current base: InfiniOps #819 at 70e50081.

Current head: c2d76051ca21774f0b515dfd8afe59537e200ac2.

Scope

This provider implements the decode contract used by InfiniLM. It intentionally rejects KV append, rotary embedding, cache remapping/left padding, local windows, softcap, LSE output, and scalar cache lengths. Unsupported forms fail closed before dispatch.

Stacked Review

This PR contains three decode-only commits on top of #819:

• ae2ce9c5 - native Moore paged decode provider
• c6de2f45 - paged decode smoke coverage
• c2d76051 - fail closed for scalar cache lengths

git range-diff maps the three hardware-tested pre-rebase commits to these current commits one-for-one. After #819 merges, this PR can be retargeted to master without changing its review diff.

Validation

Operator coverage

• D64/D128 x FP16/BF16 x MHA/GQA.
• Cache lengths 1/255/256/257, non-sequential block tables, tail -1, ALiBi and scale.
• Non-contiguous outer strides, cache immutability, empty rows, non-default streams.
• Same-pointer relaunch after in-place cache-length and block-table updates.
• Final combined Moore attention smoke: 4 passed, 8 deselected.
• Moore library and generated Python bindings built successfully.

InfiniLM integration

Validated with InfiniLM #506 and the current Moore runtime/communication prerequisites:

• 13/13 selected real-weight Moore inference commands passed.
• The matrix includes paged prefill and decode, explicit FlashAttention, segmented graph replay, TP2/TP4 BF16 collectives, and MiniCPM4 Eagle speculative decoding.
• All 13 commands reported graph segments and host_segments > 0; no command fell back to wholly eager execution.

CI

Clang-format, Ruff, and documentation checks pass. Because this is a stacked PR whose base is #819 rather than master, the repository backend matrix is not triggered yet; it will run after retargeting.

Notes for Reviewers

• This PR does not add a public deprecated attention operator. It implements the current FlashAttnWithKvcache API directly.
• Scalar cache lengths are rejected deliberately: accepting them would make graph replay observe stale metadata.
• The InfiniLM integration also needs InfiniCCL fix: add -std=c++17 to Iluvatar CUDA flags #69 for Moore BF16 collectives; that dependency is independent of this operator PR.