Implementing scalar indexing for device-side sparse arrays

[alonsoC1s]

This PR addresses #694 by implementing scalar indexing for the mentioned array types. Only scalar indexing is implemented, since doing vectorized indexing would result in dynamic allocations on the device, which is not allowed

Remaining things to be fixed:

• Indexing for GPUSparseDeviceMatrixCOO (re-implementing it the way CUDA.jl does it doesn't work for some reason)
• Logical indexing
• Scalar indexing for BSR formatted arrays
• Including initial indexing tests
• Add COO, CSR, BSR tests

[kshyatt]

Any chance the rest of the TODOs here could be addressed soon? I'd like to get this in before the next tag :)

[alonsoC1s]

@kshyatt of course! I'm working on logical indexing atm. I do need some help to get the tests added (and the checks passing), though

[kshyatt]

Will this conversion work? You might need to force i::Ti above, not sure...

[alonsoC1s]

I haven't checked with less common number types. This is probably something that would come up naturally during the full testing suite

[alonsoC1s]

I finally managed to add initial testing functionality. Notably missing are the tests for any sparse format other than CSC. This is true not only for the proposed changes, but, if I'm not mistaken, formats such as COO are never used in the constructor, linalg, etc. tests

• I changed the indexing implementation to introduce the fixes recently made in CUDA.jl CUSPARSE: fix COO scalar getindex crossing row boundaries (#3100) CUDA.jl#3101 for relevant types
• From initial testing, the conversion convert(Ti, i) poses no problems. Do note that all tests currently only change the eltype (Int{16, 32, 64}, Float{...}, Complex{...}), but the index type is always Int64