Deterministic floating-point arithmetic

[eggrobin]

• find the right set of flags to prevent floating-point contraction while still allowing optimizations that make assumptions about the rounding mode and ignore floating-point exceptions;
• use the same (correctly-rounded) libm on all platforms.