The constant model in dose_response uses a second optimizer for the same likelihood #85

Problem

tl/_dose.py maximises the same Student-t likelihood with two different optimizers.
_fit_maximum_likelihood runs Nelder-Mead over a model's parameters plus the error scale, and
_fit_maximum_likelihood_constant runs minimize_scalar(method="bounded") over the error scale alone, with a
hand-picked upper bound of std * 100 that nothing else in the file uses.

The constant model is the zero-response model once the baseline is subtracted, so it is the same objective with
an empty parameter vector. Making it a third entry in _winning_model's model set would delete the function,
the minimize_scalar import, the bounds heuristic, and the if constant is None branch in _hitcall.

Expected behaviour

One optimizer for one likelihood.

Note this is not a pure refactor: bounded Brent and Nelder-Mead will not land on bit-identical optima, so the
P1 term of every hitcall moves in the last digits. It needs the dose tests re-run and the tutorial 11 outputs
regenerated, which is why it was left out of the cleanup that introduced the dataclass.

Related

• The _Fit dataclass added in 337dbcb already carries what a "constant" entry would need: a name, a parameter
  vector whose last element is the error scale, and a predict.