Add critical preprocessing transforms for robust protection testing

[Claude]

The robustness testing suite included only 4 basic transforms (identity, gaussian blur, 2 resize operations), missing critical preprocessing steps that real-world mimicry pipelines (DreamBooth/LoRA) actually use. This created a false sense of robustness—protection may survive gaussian blur but fail against JPEG compression, center crops, or common training augmentations.

Changes

Extended transform suite from 4 to 12 transforms:

• JPEG compression (jpeg_compress_decompress) at quality 95/85/75 - most effective purification defense, uses real JPEG artifacts via PIL instead of bilinear resize proxy
• Center crop (center_crop_and_resize) at 90%/80% - standard data augmentation in training pipelines
• Random crop (random_crop_and_resize) at 90% - stochastic training augmentation
• Color jitter (color_jitter) - brightness/contrast/saturation/hue shifts common in training
• Gaussian noise (add_gaussian_noise) at σ=0.01 - regularization noise testing
• Stronger blur - added gaussian_blur_medium at σ=2.0 for parameter space exploration

Backward compatibility: Renamed original gaussian_blur → gaussian_blur_mild to clarify it's part of a spectrum.

Example

from auralock.core.style import build_style_transform_suite

suite = build_style_transform_suite()
# Returns 12 transforms including:
# - jpeg_quality_95, jpeg_quality_85, jpeg_quality_75
# - center_crop_90, center_crop_80
# - color_jitter_mild, gaussian_noise_small
# - gaussian_blur_mild, gaussian_blur_medium
# - resize_restore_75, resize_restore_50
# - identity

All transforms preserve image shape and bounds [0, 1], handle tensors with gradients (.detach() in JPEG), and include comprehensive test coverage.