PyO3: Add mkl support

[LLukas22]

Alright this should add mkl support. I also created a quick benchmark script to only test the matmul performance between torch and candle: https://gist.github.com/LLukas22/b58adc148e771afdeaebc4074f0644f7

Results without --features mkl:

PyTorch: 10.230922899999769
Candle: 3.1102483999998185

With --features mkl:

PyTorch: 10.196486699999696
Candle: 4.144704999999703

I'm running this on an AMD Ryzen 7 3700x, ans seeing these results i guess Sarah did a pretty good job optimizing the gemm create.

I also re-ran the embedding "benchmark" and candle still is a bit slower there, but faster than without mkl. I'm guessing we lose some speed in other operations than matmul or trough the python wrapper.

With mkl:

Without mkl:

That being said i'm probably gonna re-run these tests tomorrow on an intel based system and see if that changes anything.

[LaurentMazare]

Merged, thanks! Might be worth doing the same for accelerate on macos platforms too at some point.