fix: bound InfiniLM paged decode by block table capacity

[voltjia]

Summary

• Clamp split-KV decode sequence length to the token capacity reachable through the supplied block table.
• Apply the same bound to both warp and CTA split-KV decode helpers.

Motivation

Paged decode receives cache_lens plus a block table with max_num_blocks_per_seq entries. If the sequence length is larger than the tokens addressable by the table, split-KV decode can iterate beyond accessible block-table entries. The kernel should only scan tokens reachable through the provided table.

This PR keeps the change isolated to that bounds fix. It does not include the TP operator-cache fix from #738 or the random-sample cache-key fix from #739.

Closes #

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 / CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

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

clang-format -i src/native/cuda/ops/paged_attention_infinilm/kernel.cuh

git diff --check
# passed

ruff format --check and ruff check were not run because ruff is not installed in the remote host/container used for validation, and this PR does not modify Python files.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	C++ formatting and whitespace checks passed.	No standalone full-platform test was run for this isolated kernel bounds change.
Iluvatar	No	N/A	N/A
MetaX	No	N/A	N/A
Cambricon	No	N/A	N/A
Moore	No	N/A	N/A
Ascend	No	N/A	N/A

Benchmark / Performance Impact

No expected performance impact. The change only caps the scanned range when cache_lens exceeds the reachable block-table capacity.

Notes for Reviewers

• The computed bound is max_num_blocks_per_seq * page_block_size.
• Both split-KV helper variants use the bounded sequence length before computing shard ranges.