Launch sm90 attention and fused quant kernels on the current torch stream (CUDA-graph capture correctness)

[shanai13]

All kernel launches in csrc/qattn/qk_int_sv_f8_cuda_sm90.cu (2 sites) and csrc/fused/fused.cu (8 sites) use <<<grid, block[, smem]>>> with no stream argument, i.e. the legacy default stream, rather than PyTorch's current stream.

In ordinary eager use this is invisible (PyTorch's default current stream aliases the legacy stream). But under CUDA Graph capture — e.g. torch.compile(mode="reduce-overhead"), or a manual torch.cuda.graph() around a transformer that uses sageattention — the capture runs on a non-default stream, so these launches escape the capture: the graph records everything except the attention/quant kernels, and replays return stale output buffers (silent corruption, no error). In our Wan2.1-1.3B DiT the replayed forward was wrong by ~80% relative magnitude and appeared to run 2.3× faster, because the attention work was simply missing. The same defect would order incorrectly with any multi-stream pipeline.

Fix: take at::cuda::getCurrentCUDAStream() in the dispatch functions and pass it as the launch stream (plus the ATen/cuda/CUDAContext.h include). After this change, a manual CUDA-graph capture of a Wan2.1-1.3B forward with sage self-attention replays bit-identically to the direct call (verified at 5 timesteps on an H100 / sm90, with no measurable performance change in normal eager use).

The sm80/sm89 qattn files have the same pattern and could be fixed identically; this PR touches only the paths we could test on hardware (sm90 + fused).