Improve Fit class, add confidence method

[cdeil]

This PR improves the Fit class in gammapy.utils.fitting a bit, doing some polishing.
It also adds a new feature Fit.covariance as a new feature, wrapping Minuit.minos.
Seems to work OK for one simple example, see test_confidence

I also made a small change to the Parameters class, so that one can access parameters not just by name and number, but also to pass a Parameter object in.
This is what Sherpa does, users can pass in Parameter objects, not just names, see e.g. https://sherpa.readthedocs.io/en/latest/quick.html#a-single-parameter .
I'd like to evolve the Fit API more i this direction, because if you have a double-Gaussian, then you currently can't reach the second parameter by name. Maybe that globally unique name can be achieved, maybe not, but working with Parameter objects more in the API seems like a nice solution to me.

@adonath - Please review, especially if you're OK to have the sigma option in confidence, or if you want or need some other API, like being able to select a TS difference to reach instead. Did we make a decision there in the discussion on API to select confidence intervals?
Note that both MINUIT and Sherpa use an option "sigma" here:
https://sherpa.readthedocs.io/en/latest/fit/index.html#changing-the-error-bounds
So I think I'm +1 to also go that route, but I'm not sure. Especially the whole question of one vs two-sided is a bit more unintuitive with "sigma" ,no?

[adonath]

Thanks, @cdeil! I've left a few minor in-line comments. Let's do the discussion about, when to return a dict or a new Result object face to face.

[adonath]

I think this logic is at the wrong place, no? I think it should mainly reduce to the one line covar_result = self.covar(**covar_opts), the rest should be handled either in Fit.covar or the Registry. (like raising an error if a backend does not support covar). There is no need to set "covar_opts.setdefault("backend", "minuit")" either, as it is already set as a default in Fit.covar.

[cdeil]

The registry always gives an error. Here I want to emit a warning and continue.
I've changed this a bit, using a simpler if clause, which I prefer over try-except in this case.

[adonath]

Add section Parameters to the docstring?

[cdeil]

Done

[adonath]

As I said already in private communication, I'd be fine to remove this method, as it can be achieved by users with a few lines of Python.

[cdeil]

done

[adonath]

Add section Parameters to the Fit.covar docstring?

[cdeil]

done