Add MLP & QLoRA Fused Ops and Kernels, Mixtral

[fabianlim]

Completing more items in #25 .

• decided to remove the L40 benchmarks.

Verified that we can reproduce the roughly 20% speedups using fused-ops and kernels

• these are per device throughputs, so for two gpus we should multiply by 2 to get the actual througput
  

Verified that we are reproduce the 75% in memory reduction using 4bit base weights

• also with FSDP when using two gpus, we see another 50% memory reduction
  

[fabianlim]

running a set of benches now. will merge after complete

[fabianlim]

@achew010 pls update if you have obtained the new benches.