Safetensors and model sharding

[casper-hansen]

This PR introduces the following:

• sharding model weights, 10GB per shard
• saving safetensors
• loading safetensors

Resolves #36

Checks:

• Quantize model to safetensors
• Quantize model to torch weights
• Quantize model to safetensors with shards
• Quantize model to torch weights with shards
• Load safetensors without reference to quant filename AND with reference to quant filename
• Load torch weights without reference to quant filename AND with reference to quant filename
• Load safetensors from local directory and remote huggingface repository
• Load torch weights from local directory and remote huggingface repository