`inspect()` performance: batch `tolist()` and reuse `fg_mask`

[perctrix]

Problem

inspect() in mipcandy/data/inspection.py has two unnecessary performance bottlenecks:

1. Per-element tolist() in class_locations (line 294)

class_locations[class_id] = [tuple(coord.tolist()[1:]) for coord in indices]

This iterates over up to 10,000 single-row tensors, each triggering a Python-C++ bridge call. Should be a single batch operation:

class_locations[class_id] = tuple(tuple(loc) for loc in indices[:, 1:].tolist())

2. Redundant label != background (line 275 vs 301)

The boolean mask is computed twice on the full label tensor:

indices = (label != background).nonzero()   # line 275
# ...
fg_mask = label != background               # line 301 (same operation)

Should compute once and reuse.

Benchmark

Tested on PH2 (200 2D samples) and BRaTS (368 3D volumes, ~240x240x155):

PH2

Variant	Time	Speedup
baseline	31.7s	1.00x
batch tolist	16.6s	1.91x
+ reuse mask	16.7s	1.90x

BRaTS

Variant	Time	Speedup
baseline	360s	1.00x
batch tolist	341s	1.06x
+ reuse mask	335s	1.08x

On 2D datasets, opt1 alone nearly halves execution time. On 3D datasets, I/O dominates but both optimizations still provide measurable improvement.

Notes

• Streaming statistics (replacing torch.cat + np.percentile with online mean/std and reservoir sampling) was also benchmarked but showed no improvement (even slightly slower on BRaTS) and introduced approximation error in percentile values. Not worth pursuing.