fix: handle Ascend paged flash attention layout

[baominghelly]

Summary

• Convert paged K/V ACL descriptors from the public BnBsND layout to the CANN-required BnBsH layout in src/native/ascend/ops/flash_attn_varlen_func/kernel.h.
• Keep query and dense K/V sequence lengths as cumulative endpoints while passing paged K/V lengths as per-batch lengths.
• Add FP16/BF16, causal/non-causal, GQA, multi-block, and reordered-block-table coverage in tests/test_flash_attn_varlen_func.py.

Motivation

Ascend paged prefill could produce corrupted output when InfiniLM passed a nontrivial block table. CANN PagedAttention expects the cache layout and length convention used here, while InfiniOps publicly preserves its existing paged tensor contract. No issue is currently linked to this change.

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Ascend build/install:
cmake --build /workspace/build/InfiniOps --target infiniops ops -j16
cmake --install /workspace/build/InfiniOps

Ascend affected test:
27 passed, 9 skipped in 2.22s

Command:
PYTHONPATH=/workspace/prefix python -c 'import sys,types; m=types.ModuleType("infini"); m.__path__=["/workspace/prefix"]; sys.modules["infini"]=m; import pytest; raise SystemExit(pytest.main(["tests/test_flash_attn_varlen_func.py","--devices","npu","-q"]))'

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	N/A - not affected
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	No	N/A - not affected	N/A - not affected
Moore	No	N/A - not affected	N/A - not affected
Ascend	Yes	build and affected tests passed	27 passed, 9 skipped; eager ChatGLM3 paged flash-attn inference produced coherent output

Full `pytest` output (optional)

.....s.....s.....s.....s.....s.s.sss                                     [100%]
27 passed, 9 skipped, 14 warnings in 2.22s

Benchmark / Performance Impact

N/A

Notes for Reviewers

• The InfiniOps public tensor remains [block_num, block_size, kv_heads, head_dim]; only the ACL descriptor flattens the last two dimensions without copying data.
• Dense K/V still uses cumulative endpoints. Paged K/V uses each batch's actual length, matching CANN PagedAttention.
• Graph-mode corruption is tracked separately in InfiniLM and is intentionally outside this operator-layout fix.
• Final e2e log: /workspace/model-matrix-route-a-20260910/final-pr-smoke-chatglm.log on the Ascend 910C test container.