Latte Training: embed cache and distributed training

[wtomin]

This commit includes:

• [x] Training un-conditional Latte on Sky TimeLapse dataset with embedding cache (128 frames);
• [x] Mixed Precision: support (1) Float16; (2) BFloat16 (set patch_embedder to "linear");
• [x] Distributed training support.
• [x] Examples of the generated videos from the trained Latte model at 500k steps.
• [x] Flash-Attention Bug Fix in ms version 2.3.0.20240318