CUSPARSE: Better error msg for unsupported sparse mm

[Roger-luo]

currently when the element-type happens to be mix of Float32 and Float64 (which happens quite often for users when they use cu for the conversion) mul! will just call the generic fallback and if CUDA.allowscalar is not set, this would just throw a warning a while later.

This PR let this dispatch path just error since it's not supported for now anyways.

[Roger-luo]

it seems CI failure unrelated to this PR

[maleadt]

Although I like better error messages, which is why I filed https://github.com/JuliaGPU/CUDA.jl/issues/1271, I don't think that adding a lengthy implementation like this for every operation scales well. We need a simpler way to add such implementations (metaprogramming) or another way to detect the use of fallbacks. Isn't the scalar indexing error message clear enough?

[Roger-luo]

Isn't the scalar indexing error message clear enough?

I think scalar indexing is just a general error for CUDA since it doesn't say anything about why it is scalar indexing, e.g most of cases I'd like to know whether it's because something is not implemented.

Yeah, I'm not sure what would be the best way to detect this type of interface overloading automatically, so I just did it manually since I hit this one quite often...

[Roger-luo]

OK just had another idea, what if we make a macro @not_supported for this? e.g

@not_supported <method signature>

this is probably more readable?

[maleadt]

Yeah that seems like a better approach.