Fix: add layerwise as default

[S1ro1]

Fixes broken nccl weight transfer on vllm 0.20

---

Note

Medium Risk
Changes the hot weight-reload path for both filesystem and NCCL updates to use vLLM's layerwise reload APIs, which can affect runtime model correctness and stability during in-place updates. Risk is mitigated by keeping the kernel-format path intact and adding explicit MLA absorbed-weight recomputation for that path.

Overview
Switches checkpoint-format weight updates (filesystem and NCCL) to a new load_weights_checkpoint_layerwise flow that wraps model.load_weights(...) with vLLM initialize_layerwise_reload/finalize_layerwise_reload and set_current_vllm_config, replacing the prior process_weights_after_loading postprocessing.

For NCCL kernel-format transfers, removes the old generic postprocess step and instead explicitly calls update_mla_absorbed_weights after load_weights_kernel, then returns early.

^{Reviewed by Cursor Bugbot for commit 25f17df. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 718bbb8. Configure here.}