Add fused chunk-major LMCache staging

[yhl-amd]

Summary

• add native HIP chunk-major pack/unpack staging for the LMCache-compatible ATOM connector
• route LMCache store/load through fastpath=fused_chunk when OFFLOAD_NATIVE_KV_STAGING=1
• keep existing fastpath=chunk request-level segment-major path as fallback
• log connector fastpath in OFFLOAD load/save profile output
• add chunk-major layout, duplicate block id, and fused connector fastpath tests

Validation

• Host: python3 -m py_compile atom/kv_transfer/offload/*.py
• Host: python3 -m pytest tests/test_lmcache_offload_connector.py -q -> 19 passed, 22 skipped
• Docker yhl_kvoff_009: python3 -m py_compile atom/kv_transfer/offload/*.py && python3 -m pytest tests/test_lmcache_offload_connector.py -q -> 41 passed
• Docker native HIP smoke with OFFLOAD_NATIVE_KV_STAGING=1 OFFLOAD_CODEC_LAYOUT=segment_indexed -> native_kv_staging smoke ok
• git diff --check

Bench

Config: MiniMax-M2.5, TP=2, prefix cache on, chunked prefill on, MAXBATCH=16384, OFFLOAD_REQUEST_FASTPATH=0, OFFLOAD_NATIVE_STITCH=1, OFFLOAD_NATIVE_KV_STAGING=1, OFFLOAD_MIN_LOAD_TOKENS=1024.

• 64K c2/s4 fixed-source: avg TTFT 2.2243s, follow avg 1.5089s, load total avg 132.50ms, fastpath=fused_chunk=80, no failures
• 128K c2/s2 fixed-source: avg TTFT 7.5709s, follow avg 1.3424s, load total avg 144.56ms, fastpath=fused_chunk=86, no failures

Detailed results are recorded in /shared/amdgpu/home/hyi_qle/yhl/project/009-kv-off-llmcache/18_SUMMARY_009_kv_offload.md and code-review/lmcache-compatible-fused-staging-plan.md.