Add split metadata enforcement and CLI support for leak-free benchmarks

[Codex]

The benchmark flow risked data leakage because train/val/test separation was not enforced, Anti-DreamBooth’s holdout was only metadata, and reports lacked split lineage. This update introduces split manifests, enforces split membership in benchmarking, and clarifies holdout handling.

• Split infrastructure: New SplitType/SplitMetadata, random split creator, manifest save/load, overlap validation (src/auralock/benchmarks/splits.py), plus README guidance and auralock split create/validate CLI.
• Benchmark enforcement: ProtectionService.benchmark_file/benchmark_directory now require split metadata, validate membership, warn on non-test splits, and embed split info in reports; CLI benchmark mandates --split-manifest/--split-type.
• Anti-DreamBooth clarity: Notes updated to treat set_C as holdout; tests made path-agnostic.
• Tests: Added coverage for splits manifest and updated CLI/Docker benchmark tests to import correct modules.

Example:

# Create deterministic splits
auralock split create ./artworks --output splits.json --train-ratio 0.7 --val-ratio 0.15 --test-ratio 0.15

# Benchmark only declared test images; warns if not test
auralock benchmark ./artworks \
  --profiles safe,balanced \
  --split-manifest splits.json \
  --split-type test \
  --report reports/benchmark.json