Slice 3: GPU Pass 1 — batched forward DTCWT + phase extraction

[joeljose]

Context

Parent: #11 | PRD: #10 | Design: docs/design/gpu-acceleration.md Section B, Pass 1

Goal

Implement the first GPU pass: batch frames through DTCWTForward, extract phase deltas on GPU, transfer to CPU, and cumsum into cumulative phase arrays.

Acceptance Criteria

• _gpu_forward_pass(data, nlevels, biort, qshift, device) function returns phase_arrays (list of numpy arrays per level, shape (num_frames, num_coeffs), float32)
• Batch size auto-tuned from available VRAM: free_vram * 0.7 / (frame_vram_estimate)
• Phase deltas computed via vectorized conjugate multiply within each batch (not per-frame loop)
• Cross-batch boundary: last frame's normalized coefficients carried to next batch, verified bitwise-identical to single-batch
• Phase deltas accumulated via np.cumsum(axis=0) on CPU (float32)
• OOM caught with actionable error message
• Progress reporting (% frames processed)
• Standalone testable: can extract phases and verify shapes/finiteness without Pass 2

Blocked by

#13 (Slice 2) — needs --gpu flag and import infrastructure.

Files likely involved

• motion_mag.py — _gpu_forward_pass() function

Testing approach

• Unit test: forward pass on 10-frame synthetic data, verify phase array shapes and finiteness
• Unit test: batched (bs=5) produces identical phases to single batch (bs=10)
• Unit test: OOM handling with artificially small batch size

[joeljose]

Completed in PR #18, released in v2.0.0