Center cropping shifts images after odd-sized centered padding #133

Reproduced on master at ffb65074b39b67e1a686d1348f2d461c7ff3ab45.

center_crop subtracts separately rounded centers to compute its starting position. For an even-sized input and odd target, this selects the bottom/right alignment, whereas pad_to_size puts extra padding on the bottom/right. Padding then cropping can therefore lose original pixels.

import jax.numpy as jnp
import dm_pix
image = jnp.arange(9.).reshape(3, 3, 1)
padded = dm_pix.pad_to_size(image, 4, 4)
print(dm_pix.center_crop(padded, 3, 3)[..., 0])
# [[4, 5, 0], [7, 8, 0], [0, 0, 0]]

The round trip should recover the original 0–8 image. Deriving each offset from half the size difference matches centered padding and TensorFlow's crop-or-pad convention: https://www.tensorflow.org/api_docs/python/tf/image/resize_with_crop_or_pad

Prepared tests cover parity combinations, channel-first/channel-last data, batches, eager/JIT execution, exact round-trip gradients and the combined crop-or-pad API. Oversized target dimensions retain the existing no-op cropping behavior. This does not overlap the Gaussian-blur changes in PRs #119 and #131.