Refactor quantized processing functions

[sw]

To avoid code duplication when implementing additional quantization formats (#456), refactor the forward_mul_mat and forward_get_rows functions to use a table of function pointers, indexed by ggml_type.

This makes some functions non-inlined, I didn't see a regression in performance on my machine.

I tried to fix the "unused variable" warnings without complicating things too much, some are used in asserts.

[slaren]

I think this is great, considering the sizes of the rows (4096 in the smallest model), it shouldn't be an issue if these functions cannot be inlined anymore.

[sw]

@ggerganov did you want to look at this again or can we merge it?

[ggerganov]

Please don't merge yet - it's top priority for merging but I need some time to take a closer look