fix: correct NVIDIA CUDA queue guard in run_opencl_fft

[xywei]

Closes #270.

Summary

• fix an inverted queue check in run_opencl_fft for the NVIDIA CUDA path
• raise only when wait_for contains events from a different command queue
• keep same-queue events valid so the marker workaround behaves as intended

Why

The previous condition used if not evt.command_queue != queue, which is logically equivalent to evt.command_queue == queue. That raised on the safe same-queue case and failed to reject different queues.

Validation

• reproduced downstream failure in volumential on NVIDIA CUDA before the fix (RuntimeError: Different queues not supported with NVIDIA CUDA)
• after this one-line change, test/test_volume_fmm.py::test_volume_fmm_laplace passes for NVIDIA CUDA on ipa

[inducer]

Thx!