Fix FP env issues

[kripken]

1. Define only floating point exceptions we support (none). That
   avoids a lot of float pragmas we don't support.
2. Ifdef out the remaining such pragmas.

Helps #11086 , but does not fix it as user code may still be broken
due to LLVM failing ISel on those pragmas now.

[kripken]

This makes libc 164 bytes smaller.

[dschuff]

I think some change to the fenv defs might make sense (although, should we do something similar for rounding modes?), but I'm not so thrilled about ifdef-ing out the pragmas for the long term. I think we will still want to handle the constrained intrinsics in one form or another. Perhaps with an ISel fix.
This PR doesn't fix the fact that users might have code that has that pragma too, where we previously ignored it but now crash. So at least we should not mark #11086 as fixed.
But this should work to unblock the LLVM roll for now. let me test it with the new LLVM.

[dschuff]

Seems to work. I still predict that if we don't fix this for real soon, somebody will complain about the regression. Or maybe theyll just comment out their pragmas, who knows 🤷

[sbc100]

Are these designed to be disabled by simply not being defined? That seems unusual, but I don't know this part of the standard.

[kripken]

Yeah, I learned this today as well. It is quite unusual for the C spec, I agree.

Libraries may define in <fenv.h> only the macro values above they support (the others may not be defined).

http://www.cplusplus.com/reference/cfenv/FE_INEXACT/

And musl in fact checks whether things are defined in order to ifdef relevant code. Which actually is helpful in this case!

[dschuff]

Not sure if they are designed that way exactly, but musl code seems to use them that way: see e.g. https://github.com/emscripten-core/emscripten/blob/master/system/lib/libc/musl/src/fenv/__flt_rounds.c and in particular https://github.com/emscripten-core/emscripten/blob/2bca083cbbd5a4133db61fbd74d04f7feecfa907/system/lib/libc/musl/src/math/lrint.c gates its use of the FENV_ACCESS pragma on FE_INEXACT being defined.