Support general eltypes in matrix factorizations (LU, QR, Cholesky)

Matrices with eltypes other than T <: CublasFloat should be supported in matrix factorizations (LU, QR, Cholesky) by promotion to a supported eltype during the copy that the non-mutating functions have to make anyway. This already works in many cases by reusing non-mutating methods in Base.LinearAlgebra and only specializing the mutating methods; however, Base tends to promote to Float64 when cuda would be better served by Float32, so it would be better to systematically handle this within CUDA.CUSOLVER.

Already implemented for matrix division and SVD in #1453. Submitting this issue to make sure it's not forgotten for the other factorizations.