Add LMCache-compatible offload connector

[yhl-amd]

Summary

Replace the current LMCache offload worker save/load path with LMCache CacheEngine.store() / CacheEngine.retrieve() while keeping ATOM's AITER KV layout as opaque raw bytes.

Key pieces:

• add ATOMRawBytesLMCacheMetadata and ATOMLMCacheGPUConnector;
• pass the ATOM GPU connector into LMCacheEngineBuilder.get_or_create();
• add device staging APIs in ATOMKVByteCodec;
• keep scheduler-side load/save/deferred-free semantics;
• fix save-frontier accounting so LMCache/HBM-hit prefixes are not re-saved.

Why

The first LMCache-compatible version had a regression where a prefix already hit in LMCache/HBM could still be saved again with skip=0. A bad 128K case looked like:

[OFFLOAD-LOAD-SKIP] seq=5 hbm_cached=129856 lmc_cached=129792 need=-64 reason=hbm_satisfies_after_alloc
[OFFLOAD-SAVE-PROF] req=5 toks=129792 skip=0 store_ms=1442/1463ms

This PR tracks the LMCache hit save frontier and rolls it back only on load failure, so warm requests only save newly computed suffix chunks.

Validation

Host:

python3 -m py_compile atom/kv_transfer/offload/*.py
python3 -m pytest tests/test_lmcache_offload_connector.py -q
# 19 passed, 18 skipped
git diff --check

Docker yhl_kvoff_009:

cd /host_009/ATOM
python3 -m py_compile atom/kv_transfer/offload/*.py
python3 -m pytest tests/test_lmcache_offload_connector.py -q
# 37 passed

Bench Notes

Compared against previous no-fastpath segment_indexed CPU offload:

• 128K c2/s2 follow avg TTFT: 2.6139s -> 1.9139s; load avg 675.43ms -> 459.30ms.
• 64K c2/s4 follow avg TTFT: 2.3633s -> 2.2717s; load avg 478.18ms -> 457.24ms.

Server logs confirm enable_prefix_caching=True and enable_chunked_prefill=True; no failed load markers were observed.