Add configurable batch_size to CNN to fix GPU OOM on large directories

[umerkhan95]

Fixes #232

The batch size was hardcoded at 64 with no way to lower it, which caused OOM on GPUs with limited memory when processing large image directories.

Changes:

• CNN() now accepts a batch_size parameter (default 64, fully backwards compatible)
• features are moved to cpu after each batch instead of accumulating on gpu
• _collate_fn handles the case where all images in a batch fail to load
• bad_im_count now counts individual bad images, not just batches
• parallelise uses pool as a context manager to clean up workers properly
• single image path uses shape[0] == 1 check instead of squeeze()

Usage:

from imagededup.methods import CNN
cnn = CNN(batch_size=8)  # lower batch size for limited gpu memory
duplicates = cnn.find_duplicates(image_dir='path/to/images/')

Tests:

All 251 existing + new tests pass across the suite. Added tests for custom batch sizes, batch size edge cases (1, larger than dataset), and all bad images directory.