[Weight Transfer]: Support layerwise reloading for online quantization

[S1ro1]

Note

Medium Risk
Touches inference weight hot-reload and NCCL/filesystem broadcast flows; incorrect flag propagation or vLLM layerwise reload behavior could break online weight updates or degrade inference stability, especially under FP8 quantization.

Overview
Adds support for vLLM layerwise weight reloading during online inference weight updates, primarily to enable online FP8 quantization (inference.quantization=fp8_per_block).

Configuration is extended with quantization and a weight_broadcast.layerwise flag (propagated/validated across trainer/orchestrator/inference), auto-enabling layerwise reload when quantization is set and forbidding incompatible combinations (e.g., layerwise vs quantize_in_weight_transfer).

Inference weight update RPCs now carry the layerwise flag end-to-end (server → worker, and orchestrator → /init_broadcaster), and weight loading is refactored to a shared load_weights_checkpoint_or_layerwise helper that uses vLLM’s initialize_layerwise_reload/finalize_layerwise_reload plus a workaround to run FP8 online conversion without torch.compile when needed.

^{Reviewed by Cursor Bugbot for commit 3974fd9. 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 0683bba. Configure here.}

[samsja]

LGTM sir