ePAI: lossless GPU-resident export (~1.85x, no accuracy loss)

[aperson30]

What

Adds the lossless GPU-resident export path for ePAI — the biggest no-accuracy-loss speedup found by profiling.

Profiling showed ePAI's export/convert stage is ~158s of CPU resampling of the 26-channel logits back to native resolution — as expensive as the ~159s inference. Moving that resample to the GPU cuts it to ~0.3s (measured 50–60×) at 99.99% voxel agreement (boundary jitter only, no lesion-level change) → ~1.85× whole-pipeline, no accuracy cost.

How

scripts/epai_predict.py (+ .sh wrapper) reuses the fork's own export_prediction_from_logits — which produces both the segmentation and the PDAC/cyst/PNET findings CSV — but with ConfigurationManager.resampling_fn_probabilities patched to the GPU resampler. So the findings report is generated identically; only the resample runs on GPU.

Wired via the existing EPAI_SCRIPT_PATH hook in _run_epai_inference — no change to the dispatch logic.

Includes the fork-specific fix: determine_do_sep_z_and_axis returns the anisotropy axis as a bare scalar on the separate-z path (standard nnU-Net returns a 1-element list), which otherwise crashes resample_torch's assert len(axis)==1.

Before enabling (deploy gate)

1. py_compile passes ✅
2. End-to-end parity check (I'll run this): one real case through epai_predict.py vs the current bare-CLI path — confirm the segmentation and the output.csv findings match.
3. Then set EPAI_SCRIPT_PATH=.../scripts/epai_predict.sh in prod .env and restart.

The GPU-export lever itself is already validated (50–60×, 99.99% agreement on 2 cases); the parity check just confirms the CSV path is faithful before it touches users.