feat(preprocess): add inner_crop_from_probe — ViT-patch crop from probe geometry

[Garrett Bischof (gwbischof)]

Adds inner_crop_from_probe(probe, threshold=0.5) to ptychoml.preprocess.

Why

The reconstruction probe's amplitude defines which pixels of each ViT output patch carry meaningful signal. For a roughly circular probe of radius R (measured at threshold × peak amplitude), the largest inscribed axis-aligned square gives:

inner_crop = floor(patch/2 − R/√2)     # clamped to [0, patch//4]

Extracted from the geometry half of holoptycho's inner_crop_from_onnx so holoptycho can import it — the ONNX-loading / probe-tensor extraction glue stays local (filename convention, onnx lib). Accepts a complex or real 2-D probe, so it works on both a model's ONNX-baked probe and PtychoViTInference.baked_probe.

Pure numpy, array-in/scalar-out — same boundary as compute_sample_pixel_size / normalize_mosaic.

Note

The original inner_crop_from_onnx docstring formula (patch//2 − floor(R/√2)) didn't match its own code (floor of the whole expression — they differ by ±1). I kept the code's behavior (faithful to #37) and corrected the formula in the docstring.

Tests

tests/test_preprocess.py (+8): inscribed-square value, quarter-patch clamp, complex-amplitude handling, real==complex, threshold-controls-support, empty→None, 2-D requirement, and a lock against the original geometry on a gaussian probe.

pixi run --environment ci-py312 test   # 158 passed, 5 skipped

First of the two H4 PRs (holoptycho inner_crop_from_onnx will call this); the other is true vit-only mode.