`fused_copyto!` for CUDA assumes linear indexing

[charleskawczynski]

This needs to be generalized, which is perhaps why CI in #33 is not working, and perhaps why our multi-dimensional array perf results on cuda look too good to be true (we're unfairly getting coalescence).