Faster kernels for quantized matmul on cuda

[LaurentMazare]

This enables the new "dense" quantized matmul cuda kernels, these kernels are only used for prompt processing.

On a RTX 2080, this makes the prompt processing go from 73 token/s to 98 token/s though it's arguably on as mall prompt.

# After the change
`target/release-with-debug/examples/quantized --model mistral-7b-v0.1.Q4_K_S.gguf --prompt 'Building a website can be done in 10 simple steps:\nStep 1:' -n 100 --which 7b-mistral`
avx: true, neon: false, simd128: false, f16c: true
temp: 0.80 repeat-penalty: 1.10 repeat-last-n: 64
loaded 291 tensors (4.14GB) in 0.80s
model built
...

  19 prompt tokens processed: 98.22 token/s
  99 tokens generated: 63.82 token/s

# Before
  19 prompt tokens processed: 72.66 token/s
  99 tokens generated: 60.80 token/s