Split floating-point operations into four steps.

1. Load operands from register and convert it into a form suitable for host.
2. Execute operations specified by opcode.
3. Normalize Nans if host and guest architctures handled them differently.
4. Encode results as required by RISC-V (if host doesn't do that).

Bug: 265372622

Test: berberis_host_tests/berberis_host_tests

Change-Id: I5e4566577ef0d747147a5a7e26345b80fa7c8446