Quanto

New features:

- quantize weights and activations parameters
- float8 activations

New features:

- weight-only quantization,
- integer matmul acceleration on CUDA.

Bug fixes:

- actually use float16 weights,
- avoid float16 overflows,
- correct device placement,
- robust serialization.

New features:

- per-axis quantization

New features:

- support `opt` models,
- support `gpt-neox` models,
- support `codegen` models.

New features:

- support MPS devices,
- support Transformer models

Fix release to add correct package metadata.