Fix FlashRL compatibility with vllm 0.10.2 and torch 2.8

[nightlessbaron]

This PR addresses critical compatibility issues with newer VLLM versions (0.10.2+) and PyTorch 2.8, ensuring FlashRL continues to work seamlessly with updated dependencies.

Changes Made

1. VLLM Parameter Attribute Compatibility (flash_rl/vllm_patch.py)

   Issue: Parameter objects lost tp_rank and tp_size attributes after PyTorch/VLLM upgrades.

   Fix: Added tp_rank and tp_size to recorded_loader_keys list to ensure these critical distributed training attributes are preserved during model loading.

2. Memory Pool Management (flash_rl/fp8loader.py)

   Issue: VLLM memory pool API changes in newer versions

   Fix: Updated memory pool context management to work with new PyTorch memory allocation APIs

   Changes:

   • Updated imports to use new MemPool class from torch.cuda.memory
   • Refactored disable_mem_pool context manager to use new thread-local allocation APIs (_cuda_beginAllocateCurrentThreadToPool, _cuda_endAllocateToPool)
   • Added robust pool fetching with fallback handling for VLLM's internal pool storage changes

[nightlessbaron]

I believe this should works now. Can you review it @LiyuanLucasLiu ?