Re-evaluate what space adaptivity happens in with the logit transform turned on

[dmcdougall]

It appears that adaptive metropolis happens (computing sample covariance matrix) in the transformed (unbounded) space. That is, we compute the sample covariance matrix of logit-transformed samples and update the TK with this matrix.

This approach means we have some logit-transform logic that lies outside of the transition kernel and in the sequence generator. If, instead, we do the adaptive metropolis procedure in user-space (on samples that have not been logit-transformed), then one could conceivably hide all the transform logic in the transition kernel instead.