TOSA: fold cast-to-bf16(cast-to-f32(x)) -> cast-to-bf16(x)

[mgehre-amd]

If we have a sequence of casts that first go to a bigger (f32) and then to a strictly smaller type (bf16), we can fold this and avoid the intermediate bigger type (f32).

[jorickert]

Are you sure that this folding is valid in all cases?
clang and gcc both compile a similar pattern also to different IR/assembly: https://godbolt.org/z/xcYs3M9Ya
I tried to validate the transformation with alive2, but it times out: https://alive2.llvm.org/ce/z/wpa2C2

[mgehre-amd]

Are you sure that this folding is valid in all cases? clang and gcc both compile a similar pattern also to different IR/assembly: godbolt.org/z/xcYs3M9Ya I tried to validate the transformation with alive2, but it times out: alive2.llvm.org/ce/z/wpa2C2

Hm, then let me restrict this to bf16 -> f32 -> bf16, which clang thinks is the same as noop: https://godbolt.org/z/a418qje9a