Multi-GPU Fixed Static Noise + Batch Mode, Embedding Prompt in Metadata, Etc

[pftq]

The video has half the screen covered in static noise when using multi-GPU with 2 GPUs, 3/4 of the screen when using 4 GPUs, etc. This happens when seed is not specified because each GPU generates its own random seed and doesn't sync. This fixes that bug identified here:
#35

Additionally, I implemented the --batch-size to allow for batch creation of videos without reloading the model each time. There are also other quality-of-life features:

• Variety batch ("--variety-batch") option for varying up CFG/Steps between videos in a batch.
• Prompt details embedded in MP4 video metadata comments (requires ffmpeg installed: apt-get install -y ffmpeg).
• More useful filenames with CFG, Steps, seed, lora name, #GPUs, etc.
• Maintained aspect ratio when resizing to divide evenly among GPUs in multi-GPU spatial parallelization.

[pftq]

This also indirectly fixed #60 (comment)

[pftq]

Fixed one more bug in the resizing code. If you try 720x720 (540p) on 8 GPUs, it'll run into divisibility issues again. I managed to fix it by having the crop_size_list regenerated based on the diviser instead of 32, so it looks like the list/filter doesn't quite work and it's safer to just generate a new crop list.
crop_size_list = generate_crop_size_list(bucket_hw_base_size, diviser)
The full block:

            crop_size_list = generate_crop_size_list(bucket_hw_base_size, 32)
            aspect_ratios = np.array([round(float(h)/float(w), 5) for h, w in crop_size_list])
            closest_size, closest_ratio = get_closest_ratio(origin_size[1], origin_size[0], aspect_ratios, crop_size_list)

            if ulysses_degree != 1 or ring_degree != 1:
                diviser = get_sequence_parallel_world_size() * 8 * 2
                if closest_size[0] % diviser != 0 and closest_size[1] % diviser != 0:
                    crop_size_list = generate_crop_size_list(bucket_hw_base_size, diviser)
                    xdit_crop_size_list = list(filter(lambda x: x[0] % diviser == 0 or x[1] % diviser == 0, crop_size_list))
                    xdit_aspect_ratios = np.array([round(float(h)/float(w), 5) for h, w in xdit_crop_size_list])
                    xdit_closest_size, closest_ratio = get_closest_ratio(origin_size[1], origin_size[0], xdit_aspect_ratios, xdit_crop_size_list)