refactor(ops): migrate paged caching to canonical InfiniOps API

[voltjia]

Summary

• Replace the deprecated PagedCachingInfinilm dependency in the InfiniOps-backed paged-caching adapter with the canonical ReshapeAndCacheFlash API added by InfiniOps Issue/867： adjust paged_attention_prefill interface naming #883.
• Preserve the existing InfiniCore paged_caching(k_cache, v_cache, key, value, slot_mapping) public API.
• Present the existing HND cache tensors to InfiniOps as metadata-only NHD views; cache storage is not copied.
• Use kv_cache_dtype = auto and one persistent scalar scale tensor, which the current auto path ignores.

API Migration

Layer	Before	After	Alignment
InfiniCore public API	paged_caching(k_cache, v_cache, key, value, slot_mapping)	Unchanged	Existing InfiniCore contract
InfiniOps adapter	PagedCachingInfinilm::Call(key, value, slot_mapping, key_cache, value_cache)	ReshapeAndCacheFlash::Call(key, value, slot_mapping, k_scale, v_scale, kv_cache_dtype, key_cache, value_cache)	vLLM reshape_and_cache_flash, adapted to InfiniOps input/attribute/output grouping

Dependency: InfiniOps #883, merged as e733e325.

Validation

Remote environment: ssh nvidia, image accelerator-dev/nvidia:latest.

• clang-format --dry-run --Werror src/infinicore/ops/paged_caching/paged_caching_infiniops.cc: passed.
• git diff --check origin/main...HEAD: passed.
• Full infinicore_cpp_api and _infinicore build/link/install: passed.
• Installed-library ldd: libinfinicore_cpp_api.so resolves libinfiniops.so from python/infinicore/lib.
• Existing test/infinicore/ops/paged_caching.py --nvidia, restricted to its eight equal-head-size configurations: 24/24 passed across FP16, BF16, FP32, HND, and NHD physical cache layouts.

The full integration build used an InfiniOps validation tree containing open PRs #887-#891 because current InfiniCore adapters also reference the providers covered by those PRs. This change itself pins only merged InfiniOps #883.

Known Limitation

The final DeepSeek MLA case in the existing test uses key/value head sizes 576/512. It runs after all 24 standard cases and is rejected by the canonical API's equal-shape constraint. This is not introduced by the migration: deprecated PagedCachingInfinilm also asserts equal key/value and cache shapes. Extending that behavior belongs in a separate operator/API change.

[voltjia]

Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.