Some checks are failing, I probably messed up several functions.

By adding a bias we ensure that 1.999… will be displayed as 2.0 as long as we
do not ask for too many digits, and that rounding is reasonable for most values.
Note that our need for proper rounding is not as high as for IEEE floats since
we have an insane amount of digits at our disposal.

Also we now get rid of trailing zeroes when printing reals.

The name “re” came from “reciprocal” but since we have “inverse” for
matrices, I thought it would be nice to make everything consistent.

to get M_PI from the platform headers and cast it to float.

computation function.

compilation time for very little benefit and maybe even confusion.

computed on demand with static initialisation.

fully initialised.

"shader" source directory since we have "gpu".

that are not explicitly real, even with namespace mistakes.

that x + y != x, and nextafter() which behaves like the C function.

functions are now about 20% faster.

bonus, multiplication or division by a power of two will be optimised as
a shift of the exponent.

missing includes.

add the corresponding unit test.

to modify their precision at runtime.

negative zero or negative nth roots are not handled.

multiplications where appropriate.

the 16-to-32-bit refactoring.