Add KV cache quantization feature

[neilmehta24]

With this PR in mlx_lm, users can quantize the KV cache. ml-explore/mlx-examples#1075 . Quantizing the KV cache improves performance at large context size.

Unfortunately the KV cache quantization is only available for KVCache and not RotatingKVCache, so we would need to refactor our implementation to use KVCache to allow for quantization