GMX_FLOAT_EPS and GMX_DOUBLE_EPS don't follow ISO C standards
ISO C demands that EPS is the value of the least significant bit, whereas Gromacs uses half of that.
I don't know if this impacts any code, but we should change to the ISO C standard to avoid confusion.
It matters when you want to subtract one bit from 1, which might not happen with the halved value.
Modify single/double EPS to agree with ISO C
The previous Gromacs values were a factor 2 too tight. While
we're at it, we also added float modifiers to the single-precision
constants, increased the accuracy in the double max/min values,
and improved the least significant bit in one of the single-precision
values. The previous relative accuracies for the PME correction
analytical approximations where set fairly arbitrary by dividing a
tightly chose constant with GMX_REAL_EPS, and since this is now a
larger number the unit test tolerance for those corrections has been
increased slightly to avoid triggering false failures.