Extend `testvalue` with a `::Type` argument

[cscherrer]

If this is a universal requirement, can you also document somewhere, perhaps on the definition of testvalue that it should produce a point from the sample space of the measure such that evaluating logdensity on it produces a Float64?

Originally posted by @sethaxen in #137 (comment)

We should add methods to testvalue so that for any measure d,

1. logdensity(d, testvalue(d)) isa Float64
2. logdensity(d, testvalue(T, d) isa T