Fix JPEG transform for non-contiguous batches - #9615

Summary

• allow the v2 JPEG kernel to process non-contiguous tensors with multiple leading batch dimensions
• add regression coverage that compares a transposed batch layout with its contiguous equivalent

Problem

transforms.v2.JPEG documents tensor inputs as [..., C, H, W], with an arbitrary number of leading dimensions. The kernel flattened those dimensions with view(), which rejects valid non-contiguous layouts such as tensors with transposed batch axes before JPEG encoding is reached.

Fix

Use reshape() to flatten the leading dimensions. This remains a view when the layout is compatible and materializes the logical tensor order only when required, so the existing contiguous path is unchanged while valid non-contiguous inputs are supported. The decoded images are then restored to the original shape as before.

Testing

• python -m pytest test/test_transforms_v2.py::TestJPEG -q (78 passed)
• python -m pytest test/test_transforms_v2.py -q (6690 passed, 2936 skipped, 348 xfailed)
• pre-commit run --files torchvision/transforms/v2/functional/_augment.py test/test_transforms_v2.py

cc @vfdev-5