Offsetting kernel indices

[glwagner]

Increasingly we have the need to offset kernel indices; ie to perform computations over "windows" for incomplete Field indices:

Oceananigans.jl/src/AbstractOperations/computed_field.jl

Lines 84 to 92 in be00e36

	@kernel function _compute!(data, operand, index_ranges)
	i, j, k = @index(Global, NTuple)
	i′ = offset_compute_index(index_ranges[1], i)
	j′ = offset_compute_index(index_ranges[2], j)
	k′ = offset_compute_index(index_ranges[3], k)
	@inbounds data[i′, j′, k′] = operand[i′, j′, k′]
	end

or for @simone-silvestri's work to overlap communication and computation, eg #3067.

Possibly, such a feature can be supported by KernelAbstractions eg JuliaGPU/KernelAbstractions.jl#384 which would simplify our code a lot (since we won't have to pass the offsets into the kernel explicitly).