Represent distributions in MIR, refactor backend

[seantalts]

The backend is getting pretty complicated and in many places we're still relying on function name string suffices to determine things like which arguments need to be passed to which, and the logic is getting kinda crazy. I'd like to see some separate variants made that line up more closely with the platonic ideal of distributions (with their various grouped functions like _rng, _lpdf, _ccdf, etc).

[bob-carpenter]

This'd be good as the argument signatures should also be tied together.