Auto-byte quantize

[bratseth]

This adds automagic byte quantization to the HuggingFace embedder.

Maybe we should check for L2 normalization, or require normalization=true to do this?

/cc @thomasht86

[arnej27959]

In the non-normalize case it’s probably more useful to scale with 127.0/max(abs(x)) and assume angular distance.

[arnej27959]

And the normalize case is very likely to end with almost all values 0.

[arnej27959]

I was a bit too pessimistic; checking on Cohere embeddings doing (L2 norm)*127 I got:

• 23% of the resulting numbers are 0
• 86% are in the range [-4,4]
• 99% are in range [-10,10]