Allow specification of inverse mass matrix (Feature Request)

Summary:

@mitzimorris recently added the specification of the initial inverse matrix (https://github.com/stan-dev/stan/pull/2260). It would be nice to have this feature in rstan.

Description:

It could be added by adding an additional argument (init_inv_metric) to the sampling command. When metric is "diag_e" then init_inv_metric is a vector, or list of vectors, one for each chain, with length equal to the number of unconstrained params. When metric is "dense_e" then init_inv_metric is a semi-positive definite matrix of the appropriate size. When metric is "unit_e" and init_inv_metric is specified, then an error is thrown.

Alternatively, for "diag_e" the inverse metric could be specified on a per parameter basis, similar to how "inits" is used to set initial values.

I could take a stab at implementing this if the proposal seems reasonable, and no one else is already working on it.

Current Version:

v2.16.2

Start with the rstan sampling entry point and the handling of existing metric and init arguments, then compare Stan PR #2260. Define tests for diag_e, dense_e, and unit_e behavior; done means sampling accepts the specified inverse metric forms and rejects it for unit_e.