Decide how to handle arithmetic errors

[cgay]

The test suites have been failing on macOS for quite some time with these errors:

To my understanding we need to decide whether to signal errors for these arithmetic functions (slow) or return NaN et al (fast).

My $0.02: let's be fast.

I'm going to turn these tests off on macOS for now.

[housel]

There could perhaps be a with-floating-environment form that would allow changing floating exception modes and rounding modes within a local region. The default should of course be returning NaN. (RISC-V doesn't even have a floating point exception interrupt; if you want to catch FP exceptions the compiler has to insert code to test the exception flags, and last I checked LLVM didn't even have support for doing that.)