Change `bitselect`'s condition mask to always be integral

This came up during https://github.com/bytecodealliance/wasmtime/pull/5918 and the removal of the old `vselect` instruction. Currently the `bitselect` instruction uses the same type for all of its inputs and its output, but that means that for a floating-point selection the condition mask is also a floating-point value which is typically not the case. For example an `icmp` would produce an integer vector which could be used to `bitselect` the results of the comparisons.

Using an integral mask would make the NaN-canonicalization for vectors slightly simpler and additionally make it easier to re-add the optimizations removed in https://github.com/bytecodealliance/wasmtime/pull/5918 where `bitselect`-of-`fcmp` for float vectors could be optimized into `f{min,max}` where possible.