Add CheXlocalize dataset support (CheXpert official val/test + segmentation masks)

[AmlanMishra2004]

Proposal

Add a CheXlocalize_Dataset loader for CheXlocalize (Saporta et al., Nature Machine Intelligence 2022, https://doi.org/10.1038/s42256-022-00536-x): the official CheXpert validation/test images and labels, plus radiologist ground-truth segmentation masks for 10 pathologies — one of the contribution areas explicitly called out in CONTRIBUTING.md ("segmentation masks").

Why not CheX_Dataset

CheX_Dataset can't load this data as-is:

• It assumes Sex/Age/Frontal-Lateral/AP-PA columns exist. CheXlocalize's blinded official test_labels.csv omits them to prevent re-identification of the test set — loading it with CheX_Dataset raises KeyError: 'Frontal/Lateral' (confirmed against the real download).
• Even with those columns present, it infers the train/valid split by string-matching 'train'/'valid' in the CSV path, and raises a bare NotImplementedError otherwise — which also fails on CheXlocalize's test_labels.csv.

Planned implementation

A new CheXlocalize_Dataset(Dataset) class, mirroring CheX_Dataset's constructor/label conventions but:

• synthesizing safe defaults for the missing demographic/view columns instead of assuming they exist,
• extracting patient IDs/views without depending on the split name,
• normalizing CheXlocalize's path quirks (CheXpert-v1.0/ prefix, valid/ CSV prefix vs. val/ directory on disk),
• adding pathology_masks=True + segmentation_jsonpath=... support for CheXlocalize's ground-truth segmentation JSON (COCO RLE via pycocotools), following the same pattern as SIIM_Pneumothorax_Dataset.get_pathology_mask_dict.

I already have this implemented and verified (including against a real CheXlocalize download — test/val sample counts match the documented patient counts, and decoded masks match the CSV labels for spot-checked patients). Opening this issue per CONTRIBUTING.md before the PR; PR is up at #191.