Keep dots in sample names for YOLO and COCO datasets

[SajalDevX]

Fixes #642

What

yolo.build_dataset() and coco.build_coco_dataset() now build the sample name with os.path.splitext(os.path.basename(...))[0] instead of basename.split(".")[0].

Why

Cutting at the first dot gives every Roboflow-exported copy of an image (IMG_0001_jpg.rf.<hash>.jpg) the same key. Each copy then overwrote the previous row in the dataset dict and was silently left out of evaluation. splitext only strips the extension, so:

• img1.jpg → img1 (same as before, nothing changes for plain names)
• IMG_0001_jpg.rf.3f9a1c7e2b.jpg → IMG_0001_jpg.rf.3f9a1c7e2b (unique now)

How I tested it

• New test_build_dataset_keeps_samples_with_dots_in_name in tests/datasets/test_yolo.py. It reuses the existing mocked helper and also checks that a plain name still maps to img1.
• New tests/datasets/test_coco.py with a small COCO file written to tmp_path, covering four images, three of them with Roboflow-style names.
• Both new tests fail on master and pass with this change.
• Full suite via poetry install + poetry run pytest, as in CI: 83 passed, 1 skipped.
• Also ran it by hand on a Roboflow-style folder (5 images) and a COCO JSON (3 images). Before: 2 and 2 rows. After: 5 and 3.
• black --check is clean on all four files.

I used AI assistance while working on this (tracing the loaders and drafting the tests). I've gone through every line and ran all of the above myself.

Note: because sample names now keep their extra dots, predictions that were already saved for such files under the old (truncated) sample name won't be matched after this change and would need to be regenerated.