Improving ChainRules rule usage

[willtebbutt]

The @from_rrule macro which lets you make use of existing ChainRules.rrules to define new Tapir.rrule!!s is quite limited in what inputs it supports. In particular, it permits only

1. Float64s, and
2. types which are non-differentiable.

Moreover, the non-differentiable types must be bits types, and the ChainRules.rrule must return the correct type. e.g. it won't work if the ChainRules.rrule returns a Float64 gradient for an Int primal.

One narrow way to extend this would be to assume that we can just return a Tapir.NoRData whenever the primal is a non-differentiable bits type, regardless what the ChainRules.rrule returns.

Motivated by TuringLang/Bijectors.jl#319 (comment) .