Slice 4: GPU temporal filter — chunked cuFFT

[joeljose]

Context

Parent: #11 | PRD: #10 | Design: docs/design/gpu-acceleration.md Section B, Temporal Filtering

Goal

Implement GPU-accelerated temporal filtering using torch.fft with coefficient-dimension chunking to fit in VRAM.

Acceptance Criteria

• _gpu_temporal_filter(phase_arrays, magnification, width, device) modifies phase arrays in-place
• Large window (width=80, 344 samples): uses torch.fft.rfft/irfft along time axis
• Small window (width=2, 9 samples): also uses cuFFT (overhead is negligible)
• Chunk size auto-tuned: free_vram * 0.7 / (num_frames * 80 bytes FFT overhead)
• Phase modification: phase = phase0 + (phase - phase0) * magnification
• Smoothing pass: second filter with width=2
• Per-level processing (largest levels first), each level chunked independently
• OOM caught per-chunk with actionable error
• Works correctly for all 8 DTCWT levels (tiny levels may have fewer coeffs than chunk size — handle gracefully)

Blocked by

#14 (Slice 3) — needs phase arrays from forward pass to filter.

Files likely involved

• motion_mag.py — _gpu_temporal_filter() function

Testing approach

• Unit test: filter output has same shape as input, finite values
• Unit test: DC signal (constant phase) passes through unchanged
• Unit test: chunked filter produces same result as single-chunk (small data that fits in one chunk)
• Speed test: verify GPU filter is faster than CPU flattop_filter_1d on L0 data

[joeljose]

Completed in PR #18, released in v2.0.0