[train] Support LoRA-only weight syncing for Megatron backend

[SumanthRH]

Summary

Currently, for weight syncing with LoRA, SkyRL's Megatron backend will fuse the weights and sync the merged weights with vLLM replicas. SkyRL should support syncing just the LoRA weights. and enabling LoRA serving on the vLLM side.

This was previously attempted in #885 but it also requires vLLM changes.

1. Change Megatron's weight extractor to only transfer trainable, LoRA weights (this should be doable with export_adapter_weights (link) in Megatron-Bridge
2. Fix vLLM's weight loading in LRUCacheWorkerLoRAManager upstream to allow LoRA weight loading with in-memory LoRA weights
3. Upgrade SkyRL to use the vLLM release version with the above changes after 2. lands

Reference PRs:
verl: verl-project/verl#4632
Nemo-RL: NVIDIA-NeMo/RL#1797 (three PRs linked and some relevant benchmarking for LoRA only weight update code path there)

[SumanthRH]

@hao-aaron