This fixes compilation issue with C++23 mode against libstdc++.

Closes #2852

The issue was that `capture_by_value` was meant to be specialized
by plain type, e.g. `capture_by_value<std::weak_ordering>`, but
the TMP in Decomposer did not properly throw away `const` (and
`volatile`) qualifiers, before taking the value of `capture_by_value`.

Our implementation should be slightly faster, and has the
advantage of being consistent between platforms. This does not
have immediate user impact, because we currently use random_device
to generate random seed for resampling, but if we decide to change
this in the future, it is one less place to fix.

Now we use intrinsics when possible, and fallback to optimized
implementation in portable C++. The difference is about 4x when
we can use intrinsics and about 2x when we cannot.

This should speed up our Lemire's algorithm implementation nicely.

The previously used `make_unsigned` approach combined with the overload
set of `extendedMult` caused compilation issues on MacOS platform. By
forcing the selection to be one of `std::uintX_t` types we don't need
to complicate the overload set further.

Previously it could be just plain reporter name, e.g. `xml`, but
it could not specify other reporter options. This change is not
particularly useful for the built-in reporters, as it mostly comes
in handy for combining specific custom reporter with custom arguments,
and the built-in reporters do not have those.

The Core Guidelines state "A base class destructor should be either
public and virtual, or protected and non-virtual" so, hopefully, no
static analyser will complain.