modify parameters in-place in sciml_train!

[ric-cioffi]

This is more of a suggestion rather than an issue.
I'd find it much more intuitive to have sciml_train! to modify the parameters in-place, which would be consistent with Flux.train!'s behavior.

[ChrisRackauckas]

I am going the complete other direction in #139 and just renaming it sciml_train instead of sciml_train!. One of the purposes of this was to get rid of the global behavior because it's brittle, not thread-safe, and can lead to some weird performance issues.

[ric-cioffi]

I'm totally fine with this as well, but it should probably be highlighted in the docs that DiffEqFlux.sciml_train behaves differently from Flux.train! even when called with a Flux optimizer.

[ChrisRackauckas]

Yes, it does behave differently from Flux.train! in more than one way, another being after #139 the boolean in the callback. Another is the maxiters stuff. So I think we just break the link and don't look back. People can still use Flux.train! if they want, but we can just give an interface that's better for the kinds of applications we see.

[ChrisRackauckas]

Went the route of #139