Add low VRAM mode, CPU-only mode + image pre-loading fix

[FlyingFathead]

1. Low VRAM mode implemented: Introduced the unimplemented option lingering around in streamlit_helpers.py to run the model with half-precision (float16). Allows efficient usage on GPUs with limited memory, lowering memory requirements by half. I've tested this on a RTX 3060 w/ 12GB VRAM.

2. CPU-only mode: added an option to run the model on CPUs for users without the necessary CUDA GPU hardware, though at slower speeds.

Both can be set with True/False in streamlit_helpers.py, although a more viable solution in the future would be to i.e. add a separate config.ini w/ configparser for the entire framework.

3. Image Loading Fix: Fixed an issue where the app would throw an error if an image was not loaded before model execution. The model now uses a dummy tensor until the image is loaded in. See: video_sampling.py

[linlang171]

PS D:\AI\stable video diffusion\generative-models> python scripts/sampling/simple_video_sample_4d.py
 --input_path assets/sv4d_videos/test_video1.mp4
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.0.1+cu118 with CUDA 1108 (you have 2.0.1+cpu)
    Python  3.10.11 (you have 3.10.0)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)  
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
Reading assets/sv4d_videos/test_video1.mp4
preprocess assets/sv4d_videos/test_video1.mp4
Traceback (most recent call last):
  File "D:\AI\stable video diffusion\generative-models\scripts\sampling\simple_video_sample_4d.py", line 257, in <module>
    Fire(sample)
  File "D:\AI\stable video diffusion\generative-models\.pt2\lib\site-packages\fire\core.py", line 135, in Fire
    component_trace = _Fire(component, args, parsed_flag_args, context, name)
  File "D:\AI\stable video diffusion\generative-models\.pt2\lib\site-packages\fire\core.py", line 468, in _Fire
    component, remaining_args = _CallAndUpdateTrace(
  File "D:\AI\stable video diffusion\generative-models\.pt2\lib\site-packages\fire\core.py", line 684, in _CallAndUpdateTrace
    component = fn(*varargs, **kwargs)
  File "D:\AI\stable video diffusion\generative-models\scripts\sampling\simple_video_sample_4d.py", line 106, in sample
    images_v0 = read_video(processed_input_path, n_frames=n_frames, device=device)
  File "D:\AI\stable video diffusion\generative-models\scripts\demo\sv4d_helpers.py", line 163, in read_video
    image = ToTensor()(image).unsqueeze(0).to(device)
  File "D:\AI\stable video diffusion\generative-models\.pt2\lib\site-packages\torch\cuda\__init__.py", line 239, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

I use your branch and set USE_CUDA = False.I still get this error.What did I do wrong?Maybe I need to manually modify all .to(device) to .to('cpu)?