Support `cast` to/from `i64` and `fp32`

[mtrofin]

This may be an instance of the general "(optional?) support for i64".

[mtrofin]

@banach-space @ilovepi @ioghiban @Jaddyen

[banach-space]

There are Arm CPU instructions that provide this functionality, e.g. FCVTZS from Arm's SVE ISA. I guess that it makes sense to model this at the TOSA level?

[eric-k256]

Thanks for bringing this up @mtrofin. The next release of TOSA will have INT64 as an extension, but it doesn't currently support this particular set of casts. I'm curious how these are used, is the i64 generally a tensor coordinate of some type that you then take into fp32, do some calculations in fp, and then cast back to integer?
We've got some discussion going internally on what the best way to add this would be, and will try to get an update soon.

[mtrofin]

I attached the MLIR that showcases our scenario, and here is how I got it - hope this helps!

model_tosa.mlir.zip