Manually, outline error path.

[vchuravy]

We should perhaps adapt @checked from GPUToolbox, but more importantly we likely see a lot of unnecessary small allocation due to JuliaParallel/MPI.jl#939 (comment)

CUDA.jl explicitly outlines the error path as well.

[luraess]

What's the context of these small allocs happening?

[vchuravy]

@luraess it basically means that every ccall we do to thec-API has extra overheads due to the error checking. I am not convinced it causes small allocation anymore, but we do setup a GC-frame...