[documentation] Add an optimized linear operator for sparse matrix-vector product on Nvidia GPUs

[amontoison]

For sparse matrix-vector products on Nvidia GPUs, we need to allocate some buffers.
The current implementation allocates a new buffer at each product A * v.
We could reuse the same buffer if we implement a linear operator based on the low-level CUDA wrappers.

[amontoison]

I did it in KrylovPreconditioners.jl.
The user just needs to do:

using KrylovPreconditioners
opA = KrylovOperator(A)

A can be a sparse COO, CSR or CSC matrices.
It also works for AMD GPUs.
I should add a section about it in Performance Tips or some remarks in the section GPU.