Add support for 24GB VRAM fine tuning via 8bit optimizers

[youngmae]

Caveats: Open to feedback on configuration - wasn't entirely clear how to separate it, but made it generic enough to pull in any optimizer from bnb.

Summary of changes

• Introduces dependency on bitsandbytes
• Introduces an optimizer flag to configure bnb usage
• Documentation

Tests done
5k test wav dataset, stock model config, and stock pretrained model:
| Name | Type | Params

0 | diffusion | ConditionedDiffusionModelWrapper | 1.2 B
1 | diffusion_ema | EMA | 1.1 B
2 | losses | MultiLoss | 0

1.1 B Trainable params
1.2 B Non-trainable params
2.3 B Total params
9,080.665 Total estimated model params size (MB)

With VRAM usage observed:
Device 0 [NVIDIA GeForce RTX 3090] PCIe GEN 4@16x RX: 47.85 MiB/s TX: 7.812 MiB/s
GPU 1635MHz MEM 9501MHz TEMP 77°C FAN 97% POW 315 / 350 W
GPU[|||||||||||||||||||||||||||||||100%] MEM[||||||||||||||||||22.367Gi/24.000Gi]

[rsxdalv]

Great work! Question - is it possible to make it an optional dependency? To divide the training and inference.

Edit: this PR seems to address just that https://github.com/Stability-AI/stable-audio-tools/pull/139/files