cranelift: Disable NaN Canonicalization for RISC-V

👋 Hey,

In RISC-V almost all [floating point instructions return a canonical NaN when its result would be a NaN](https://github.com/riscv/riscv-isa-manual/blob/c8155afea51a0d47297936aa2f5d29d7ce3946f3/src/f-st-ext.adoc#nan-generation-and-propagation). In cranelift we have a separate pass that includes some NaN normalization code when the `enable_nan_canonicalization` flag is enabled. RISC-V's canonical NaN pattern is also defined to be the same as WASM's canonical NaN.

This isn't really required for RISC-V and just causes additional overhead, so it would be nice if we could disable it. This thought occurred to me while reading #7321 which might make that easier by querying each backend and applying those legalization rules conditionally.