Allow per-example atol, rtol; support Ellipsis in numeric values

[ev-br]

Consider the following doctest example

>>> np.sqrt(2)
1.4...

Under doctest.ELLIPSIS this means that ... matches any string --- essentially this defines the tolerance for this particular example.

We do have doctest.ELLIPSIS on by default, and it works fine --- but it curiously breaks down under numpy 2.0 scalar representation:

>>> import numpy as np
>>> np.sqrt(2)
np.float64(1.4142135623730951)

Normally, 1.4142 would evaluate to np.allclose to np.float64(1.4142135) --- but the ellipsis breaks it because the vanilla doctest does not understand np.float64 and our machinery does not understand ....

The solution is to probably

• detect the trailing ellipsis in want, i.e. if want.edswith('...')
• chop it off
• adjust the atol, rtol to the number of sig figs (DTChecker._do_check should get explicit atol, rtol arguments defaulting to self.atol, self.rtol).

[lesteve]

I think I bumped into this in the context of scikit-learn/scikit-learn#30496 and kind of expected the ellipsis to work with dt_config.strict_check = False but at the same time, I figured "fair enough ellipsis are the old way of doing it and I should not expect floating point comparison in this case". It was straightforward enough to update the doctests to use Numpy 2 formatting in scikit-learn/scikit-learn#30495.