By moving to use our `uniform_integer_distribution`, which is
reproducible across different platforms, instead of the stdlib
one which is not, we can provide reproducible results for `float`s
and `double`s. Still no reproducibility for `long double`s, because
those are too different across different platforms.

* Utility for extended mult n x n bits -> 2n bits
* Utility to adapt output from URBG to target (unsigned) integral
  type
* Utility to reorder signed values into unsigned type while keeping
  the order.

Specifically we add
 * `gamma(a, b)`, which returns the magnitude of largest 1-ULP
   step in range [a, b].
 * `count_equidistant_float(a, b, distance)`, which returns the
   number of equi-distant floats in range [a, b].

Co-authored-by: Martin Hořeňovský <martin.horenovsky@gmail.com>

Technically, the declaration should not have a space between
the quotes and the underscore, because `_foo` is a reserved
identifier, but `""_foo` is not. In general it works, but newer
Clang versions warn about this, because WG21 wants to deprecate
and later remove this form completely.

The basic idea was to reduce the number of things dependent on the `Clock`
type. To that end, I replaced `Duration<Clock>` with `IDuration` typedef
for `std::nanoseconds`, and `FloatDuration<Clock>` with `FDuration`
typedef for `Duration<double, std::nano>`. We can generally assume that
any clock's duration can be expressed in nanoseconds, as long as we insert
`duration_cast`s into the right places.

Note that we cannot remove all dependence on `Clock` as a template
arguments, because functions that actually measure the elapsed time have
to use the Clock.

We also changed some template function arguments to pass plain function
pointers, so that the actual implementation can be placed into a cpp file.