Implement "voxelized" tokenizer

[kjappelbaum]

in https://arxiv.org/pdf/2305.05708.pdf

[kjappelbaum]

in Meta's paper (/cc @smiret-intel)

[kjappelbaum]

For building the tokenizer we can do two routes:

• Observe the extremes of the coordinates and then add everything between the extremes
  • This has an interplay with the resolution
  • This is what the paper from the Aspuru-Guzik group did
  • This might be easier with fractional coordinates
• Fit on some observed data and then only put those in the vocab

the second approach will limit generalizability, the first will give a very large vocab. Are there any other things that come to mind that we should consider, @smiret-intel , @n0w0f ?

[n0w0f]

I am lookin at Regression Transfomer tokenizer implementation in this branch.

Pros:

• Smaller Vocab
• no resolution issues

Cons:

• Requires Pretraining ? ( Not a similar treatment of number, as seen in the pretraining corpus by big models)