[Feature] KV offload: hybrid bundle backend + dense/hybrid split

[yhl-amd]

Summary

Adds a second KV-offload backend and reorganizes atom/kv_transfer/offload/
into two clearly-separated backends over a shared worker-side base.

• hybrid/ — profile-driven bundle offload: one opaque offload unit per
  128-aligned terminal checkpoint. Targets hybrid-KV models (DSV4 CSA/SWA, Qwen3-Next).
  Adding a new hybrid-KV target = a new HybridProfile, not a new connector.
  Components: connector, admission (state-pool + max-inflight-saves gate),
  store, kv_bundle + codec, policy, and profiles/ (dsv4, qwen3_next).
• dense/ — the existing token-chunked MHA/MLA LMCache offload path, moved out
  of the top-level connector.py (DenseOffloadConnector + GPU connector + byte codec).
• _offload_common.py — worker-side plumbing shared by both backends:
  separate save/load copy executors (load never queues behind fire-and-forget saves),
  lock-guarded completion tallies, the KVConnectorOutput reporting contract, and the
  opaque-uint8 LMCache engine build (ATOMRawBytesLMCacheMetadata + KV_2LTD hack).
• triton_offload_gather.py — Triton gather/scatter for staging compute-slot state
  into pool slots on the RPC thread before the next forward overwrites it.
• Top-level connector.py is now a thin dispatch shell (927 → mostly removed);
  scheduler.py gets the small hooks (should_defer_free / save-frontier) both
  backends rely on.

Test plan

• tests/test_hybrid_offload_* — admission, connector (GPU), gather, kv_bundle,
  kv_bundle codec, policy
• tests/test_offload_profile_sources.py, tests/test_offload_qwen3_next.py,
  tests/test_offload_shell.py
• tests/test_lmcache_offload_connector.py (dense path) updated for new layout
• TP4 real-run smoke on DSV4 (GPU0-3) — save/load round-trip via serve

Notes

Squashes two commits: the DSV4 terminal-checkpoint offload unit and the
subsequent dense/hybrid split. No changes to the dense wire format; hybrid is
additive and gated behind its own connector/profile selection.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 1683 --add-label <label>