Minor optimization of FFTGrid initialization

[abussy]

With PR #1056 and #1061, the bottleneck of PlaneWaveBasis instantiation becomes the creation of r_vectors at FFTGrid initialization. For systems with high cutoffs, this is non-negligible.

This PR enables the calculation of the r_vectors on the GPU. Additionally, it greatly speeds-up the CPU version by precomputing and recasting the inverse of fft_size.

[mfherbst]

Looking at the diff of this PR again, I wonder ... what does this PR actually do ? I mean is it just taking out the fac from the inner loop and that's it ?

[abussy]

Not exactly. My tests have shown that simply taking out fac from the original construct, i.e.

VT = value_type(T)
fac = 1 ./ VT.(fft_size)
r_vectors = [(Vec3{VT}(idx.I) .- (1, 1, 1)) .* fac for idx in CartesianIndices(fft_size)]

is considerably slower than the proposed change (factor ~5x). I am not sure why that is, but I suspect it has to do with with creating a Vec3{VT} as the last step. In fact, the following is as efficient as the map construct I initially proposed:

fac = 1 ./ VT.(fft_size)
r_vectors = [Vec3{VT}(fac .* (idx.I .- (1, 1, 1))) for idx in CartesianIndices(fft_size)]

I am happy to switch to the above, if that keeps the code clearer

[mfherbst]

In my opinion it does, yes.