feat(grpo): asynchronous weight synchronization with vLLM background streams

[RUFFY-369]

Summary

This PR implements Asynchronous Weight Synchronization for the GRPO trainer. It introduces a non-blocking communication layer that allows model weights to be pushed to the inference servers (vLLM/SGLang) in the background, significantly reducing pipeline "bubbles" and increasing overall training throughput.

Technical Context

Standard weight synchronization is a blocking operation that stalls the training loop while data is sent over the network. For large models, this synchronization time can account for a significant portion of the total step time.

This implementation introduces an asynchronous synchronization worker that:

• Manages a background connection pool to the inference actors.
• Offloads weight broadcasting to a separate NCCL stream, allowing it to overlap with the backward pass.
• Implements a "wait-free" synchronization protocol to ensure the trainer never stalls due to a single slow inference worker.

Key Changes

• torchtitan/grpo/sglang_handling.py: Implemented the SGLangHandler with background connection pooling and async sync logic.
• torchtitan/grpo_train.py: Modified weight synchronization to use background streams and injected wait-free logic into the step boundary.
• torchtitan/config/job_config.py: Added async_weight_update toggle and configurable synchronization timeouts.

Modernization & Compatibility

To support modern hardware and the latest PyTorch standards, this PR includes foundational modernization for PyTorch 2.5.1+.

• Backward Compatible: Uses try...except and version guards to remain fully compatible with the existing PyTorch 2.3/2.4 baseline in the dev-updated-again fork.
• Stream Management: Uses the latest PyTorch distributed stream APIs to ensure safe overlap between compute and communication.

Verification Results (vast.ai)

• Hardware Profile: Verified on a vast.ai cluster with 2x RTX 3090 GPUs (24GB VRAM).
• Scale: Measured a 25-40% increase in training throughput compared to synchronous weight synchronization.
• Tests: Successfully ran scripts/verify_grpo_2gpu.sh, confirming that weights are correctly synchronized without race conditions.
• Cluster Stability: Verified that background NCCL streams do not contend with main training kernels on consumer-grade high-memory hardware.