This removes the __detail::_Shift class template, replacing it with a
constexpr function template __pow2m1. Instead of using the _Mod class
template to calculate a modulus just perform a bitwise AND with the
result of __pow2m1. This works because the places that change all
perform a modulus operation with a power of two, x mod 2^w, which can be
replaced with x & (2^w - 1).

The _Mod class template is still needed by linear_congruential_engine
which needs to calculate (a * x + c) % m without overflow.

I'm not committing this yet, please review and check I've not broken
anything.