Stop making assumptions about subnormals and NaN payloads

The behavior of subnormal numbers and signed zeros in particular is target-specific. Some platforms simply cannot support subnormals or signed zeros without software support code that might not be present. The only long-term solution I can think of is for LLVM to stop making assumptions about this:

- Treat operations on subnormal numbers as opaque.
- Do not constant-fold operations that produce subnormal numbers as intermediates (even if the final result is normal).
- Do not assume that the payload of NaNs is preserved.
- Do not assume that the sign of zero is preserved.

This will avoid problems on platforms that omit full IEEE 754 conformance in the interest of speed, cost, or power consumption.