[knowrohit]

Summary

• validate that minibatch is a positive integer before modulo arithmetic and kernel launch
• preserve the corrected per-block FFT geometry using BS=minibatch and grid_x=Batch // minibatch
• add CUDA regression coverage for valid and invalid minibatch values

Validation

• Managed CUDA checks passed for valid minibatches 1 and 2, including shape, dtype, and numerical equality with torch.fft.fft
• Managed CUDA checks passed for invalid minibatches 0, -1, and 3
• Baseline reproduced ZeroDivisionError for minibatch=0; the repaired equivalent command passes when catching the expected ValueError
• git diff --check passed

The local environment does not have pytest installed, so the focused pytest file could not be run directly.

Fixes #102