Preprocess sliding window

[ATATC]

BRaTS Results Summary (136 MB volumes, 4 modalities)

Method	Time	Speedup
Blosc2 mmap patch (nnUNet)	1.67s	4.55x
NIfTI + MONAI Thread	2.11s	3.61x
NIfTI read + crop patch	6.07s	1.25x
NIfTI + PyTorch (workers=0)	7.61s	1.00x
Blosc2 full read	20.17s	0.38x
NIfTI + PyTorch (workers=4)	51.09s	0.15x

1. MIPCandy scenario (full volume): MONAI ThreadDataLoader is optimal - 3.61x faster than baseline
2. nnUNet's real advantage: Not the data loader itself, but the ability to read only patches (8 MB) instead of full volumes (136 MB) via mmap - that's 17x less data per sample
3. PyTorch multiprocess is counterproductive: Pickle serialization of 136 MB tensors creates massive IPC overhead (0.15x = 6.7x slower)
4. Blosc2 full read is slow: Decompression overhead makes it slower than raw NIfTI for full volumes
5. The mmap advantage: nnUNet preprocesses data to blosc2 format, then uses mmap to read only the required patch region directly - achieving 3.6x speedup over NIfTI read+crop

This confirms that for MIPCandy's current architecture (full volume loading), MONAI ThreadDataLoader is the correct solution.

Originally posted by @perctrix in #133