More robust unified fitting in Minuit.migrad and Minuit.mnprofile (#1009

)

Closes #1008

This patch was developed in collaboration with
[ikrommyd](https://github.com/ikrommyd).

- Minuit.migrad: When we have to iterate, MnSimplex is now called by
default in between interations, since this seems to improve convergence
in pathological cases. The new keyword `use_simplex` can be used to
disable this feature. It is only used if the algorithm has to iterate,
so for well-behaved functions there is no penalty of extra function
calls. Also, when we have to iterate, the strategy is now automatically
increased to 2, since this also seems to help.
- Minuit.mnprofile: Now uses the same iterative robust fitting that
Minuit.migrad uses. New keywords `ncall`, `iterate`, `use_simplex` allow
to customize the fits. Strategy 0 is used first, and if that fails, the
code switches to Strategy 2 in the iterations.
- Minuit.mncontour: Now uses the same iterative robust fitting that
Minuit.mnprofile uses if `experimental` is True. New keywords `ncall`,
`iterate`, `use_simplex` allow to customize the fits.
- A new tutorial illustrates the benefits of iterating and strategy on a pathological function. It is also shown that Scipy minimizers perform worse than Minuit when the function is pathological.

doc/tutorials.rst

@@ -13,6 +13,7 @@ Important for most users are only the first two entries.
 
     notebooks/basic
     notebooks/cost_functions
+    notebooks/unstable_fit
     notebooks/correlated_data
     notebooks/error_bands
     notebooks/interactive