fix(hicache): handle tensor device buffers

[AnkitNakhawa]

Motivation

Fixes #42672.

NPU HiCache without FIA exposes k_buffer and v_buffer as multi-layer tensors. MHATokenToKVPoolHost used getattr(..., None) or () before iterating them, which evaluates the tensor as a boolean and raises RuntimeError: Boolean value of Tensor with more than one value is ambiguous during host-pool construction.

Modifications

• Add a small iterator that handles missing buffers explicitly without boolean evaluation.
• Preserve existing list/tuple behavior and let multi-layer tensors yield zero-copy per-layer views.
• Add a CPU regression test that constructs an MHA host pool from stacked multi-layer tensor buffers.

Accuracy Tests

Not applicable. This changes buffer discovery during HiCache initialization and does not change model computation or outputs.

Speed Tests and Profiling

Not applicable. The change runs only during host-pool construction.

Testing

• pre-commit run --files python/sglang/srt/mem_cache/pool_host/mha.py test/registered/unit/mem_cache/test_mem_pool_host.py
• python -m py_compile python/sglang/srt/mem_cache/pool_host/mha.py test/registered/unit/mem_cache/test_mem_pool_host.py
• Direct CPU assertions with PyTorch 2.14 for tensor, list, missing, and None buffer representations

The registered test was not collected locally because the minimal macOS environment does not contain the full SGLang development dependency set. Ascend NPU end-to-end validation was not run locally and should be covered by NPU CI.

Checklist

• Format code with pre-commit.
• Add focused regression coverage.
• Documentation update is not required for this bug fix.
• Accuracy and speed benchmarks are not applicable.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #38449365184
Latest PR Test (Extra): ❌ Run #38449365223
Latest PR Test (AMD ROCm 10): ❌ Run #38449365306

[OrangeRedeng]

/tag-and-rerun-ci

[ping1jing2]

/rerun-failed-ci

[vlserov]

Original issue has been fixed with PR #43711. Can be withdrawn