Cleaner & faster doctests

[jorenham]

• Remove the +ELLIPSIS workarounds when comparing rounded floats, by using FLOAT_CMP from pytest-doctestplus instead.
• Improved lmo.l_loc docs and examples, and mention that the trimmed L-location generalized not only the mean, but also unifies it with the median, minimum and maximum (which hasn't been pointed out in the literature before, to my knowledge)
• Significant speedup of the theoretical.l_comoment_from_pdf doctest, by replacing the marginal CDF scipy.stats.t.cdf with scipy.special.stdtr, so that it isn't a major CI performance bottleneck anymore.
• Fall back to the order-statistics method for weight calculation if $r+s+t>24$, avoiding (explosive) numerical errors with large moment- or trim- orders (e.g. when using trimmed L-loc for median, min or max).