Remove NUMERICFLAGS

[Algunenano]

While looking at other flags, I noticed the existence of NUMERICFLAGS which activates -ffloat-store if available.

I have 2 issues with this:

• It's only used in some parts (liblwgeom/ and libpgcommon/), while other folders know nothing about it.
• It's bad for precision (you voluntarily accept less precision) and, I guess, performance since you are not using the registers.

As of today, I get the same results for clang (without that option) and with GCC (which has that flag available), so I suggest dropping it altogether and in the case that we find some difference in an output, try to change it to allow both 80bit and 64bit precision in double operations.

[komzpa]

It became obsolete since we adopted C99 which implies -fexcess-precision=standard, which has the same executive effect as --ffloat-store but does not define how to do it ("by writing to registers"). I +1 to remove NUMERICFLAGS.

If you want to race for raw computation speed try making -ffast-math pass tests once again - it did for 2.5.

[Raúl Marín <git@…>]

In 92fc725/git:

Remove NUMERICFLAGS

Closes #4629