CUSOLVER improvements

[maleadt]

Adds ldiv! for QR from #1192, and implements lu fixing #1193.
cc @ChrisRackauckas

[ChrisRackauckas]

What about the ldiv! overloads for LU? Will those work as well?

[maleadt]

Not straight away. Looking into this, we'll have to duplicate all of LinearAlgebra.LU again into a CUDA-specific CUDA.CuLU, just like we had to do for LinearAlgebra.QR...

@andreasnoack would it be possible (without requiring a Julia 2.0-type API break) to make the Factorization classes GPU compatible using additional typevars for the pivot indices/coefficients/...? Those are currently all hard-coded to Array, which breaks GPU compatibility.

[andreasnoack]

It's a shame that we didn't finish that process before the Julia 1.0 release. Unfortunately, I'm not sure we'd be able to make that change without breaking things. We could try it and run PkgEval but I'd think that we'd see several errors. I have a vague memory of similar changes resulting in downstream failures in the past.

[maleadt]

Part of #1449