Quantized SDPA

[barronalex]

First pass at adapting @angeloskath's flash attention to support quantized keys and values.

Still needs some optimization work since it's currently faster to write out the quantized_matmuls rather than use this fused version.

E.g. 4 bit on M2 Ultra for L=32768:

Timing sdpa ... 2.51938 msec
Timing quant_sdpa ... 0.97137 msec
Timing attention ... 1.31419 msec
Timing quant_attention ... 0.92342 msec

[bghira]

jfyi i have working int8 and int4 quantised attn, MIT licensed.