Replace device-specific CUDA/XPU graph APIs with unified torch.accelerator.Graph

[Copilot]

PyTorch now provides torch.accelerator.Graph as a device-agnostic replacement for torch.cuda.CUDAGraph/torch.xpu.XPUGraph. This PR migrates all graph capture/replay code to the new API, eliminating the XPU monkey-patching shim.

Key API difference

# Old pattern (CUDA-specific)
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph, pool=pool, stream=stream):
    output = model(inputs)

# New pattern (device-agnostic)
graph = torch.accelerator.Graph(pool=pool, capture_error_mode="global")
with torch.cuda.stream(stream):
    graph.capture_begin()
    output = model(inputs)
    graph.capture_end()

Notable: pool and capture_error_mode move from capture_begin()/context-manager args to the constructor. capture_error_mode="global" is passed explicitly to preserve CUDA backward compatibility (old default was "global"; new default is "default").

Changes

• vllm/compilation/cuda_graph.py — replace CUDAGraph() + torch.cuda.graph() CM with accelerator.Graph; update CUDAGraphEntry.cudagraph type annotation to Any
• vllm/compilation/breakable_cudagraph.py — replace CUDAGraph() + conditional capture_begin(pool=...) with accelerator.Graph(pool=...) + unconditional capture_begin(); update type annotation
• vllm/v1/worker/gpu/cudagraph_utils.py — same refactoring; update graphs dict type annotation
• vllm/v1/worker/encoder_cudagraph.py — same refactoring; update BudgetGraphMetadata.graph type annotation
• vllm/v1/worker/gpu_ubatch_wrapper.py — same refactoring; update CUDAGraphMetaData.cudagraph type annotation
• vllm/v1/worker/xpu_model_runner.py — remove the torch.cuda.graph/CUDAGraph/graph_pool_handle monkey-patching lines from _torch_cuda_wrapper() (and the now-unused supports_xpu_graph import)
• tests/v1/cudagraph/test_cudagraph_dispatch.py — update mock targets from torch.cuda.graph to torch.accelerator.Graph