feat: add pre-tokenized dataset cache to accelerate dynamics training

[tashapais]

Summary

• Adds scripts/preprocess_tokens.py: one-time script that runs a trained VideoTokenizer over a full dataset and saves token indices as [N, P] int32 to HDF5. Stores metadata (latent_dim, num_bins, codebook_size) as HDF5 attrs.
• Adds TokenizedVideoDataset to datasets/datasets.py: loads the pre-tokenized HDF5 and returns [T, P] index sequences, matching the (tokens, 0) interface of VideoHDF5Dataset.
• Updates train_dynamics.py: when cached_tokens_path is set in config and the file exists, the dataloader returns token indices directly. The video tokenizer forward pass is skipped each training step, eliminating the per-batch tokenization overhead.
• Adds cached_tokens_path: Optional[str] = None to DynamicsConfig.

Why it matters: dynamics training repeats the same tokenization pass every step. With a large dataset and a GPU-bound tokenizer, this can cut 30-50% off wall-clock training time.

Usage

# 1. pre-tokenize (one time)
python scripts/preprocess_tokens.py \
    --video_tokenizer_path runs/.../video_tokenizer \
    --dataset PONG \
    --output_path data/pong_tokens.h5 \
    --device cuda

# 2. set in configs/dynamics.yaml
# cached_tokens_path: data/pong_tokens.h5

# 3. train as usual — tokenizer forward pass is skipped
python scripts/train_dynamics.py --config configs/dynamics.yaml

Test plan

• Run preprocess_tokens.py on a small dataset and verify HDF5 is created with correct shape [N, P]
• Train dynamics with cached_tokens_path set and verify loss curve matches raw-frame training
• Verify cached_tokens_path: null (default) falls back to the original frame-loading path