Expose TSNE.n_iter_ fitted attribute

[jcrist]

This:

• Plumbs through n_iter_ as a fitted attribute of TSNE. This should be an integer between 0 and n_iter (the maximum number of iterations supported), and represents the number of iterations the algorithm actually ran for.
• Exposes TSNE.n_iter_ through cuml.accel.

A couple notes:

• Our C++ APIs are inconsistent in how we handle output parameters. I went for the smallest amount of code change possible, but no strong thoughts if this is the best way to structure things 🤷.
• I'm unable to find a case where we exit early, even though all algorithms (excluding barnes_hut) do have an early stopping condition. I still think exposing this this way is good (rather than just assuming max iterations and doing self.n_iter_ = self.n_iter) since it leaves plumbing in place in case the early stopping criteria changes (or there are cases where we do succeed in stopping early).
• This only makes the test for the existence of TSNE.n_iter_ pass, since the other failing tests around n_iter_ were also related to the exact logging message shown to the user. I think that's fine.

Fixes #6857

[betatim]

Looks reasonable to me. I think having the plumping here is sensible, even if it doesn't seem needed.

My one review comment is more for my education/asking why we have all this templating if the types are anyway fixed (YAGNI)

[viclafargue]

LGTM! Just one comment.

[jcrist]

/merge