Improve Quantization

The quantization support I've added through `--low-prec-bytes-per-val` is a bit barebones. It'd be nice to add enough flexibility to handle per-block quantization (e.g. some only quantize the linears to int4) and some of the new formats that aren't a multiple of a byte (e.g. int4, fp6, etc)

Relevant: https://github.com/EleutherAI/cookbook/issues/36

Start by locating the implementation and tests for the `--low-prec-bytes-per-val` option, then read the related cookbook issue 36 for context. Clarify the quantization design and scope before changing code. Done means per-block quantization can target selected layers and formats such as int4 and fp6 are supported, with tests covering those cases.

Written by the indexing model from the issue text.