Fix for #6931: isotonic y_min/y_max bounds

[mjbommar]

As noted on #6921, the isotonic_regression method:

• assumed pre-sorted data
• relied on a heuristic reweighting to achieve the desired y_min/y_max outcome (that does not work w.p.1)

This PR fixes both issues by:

• adding a test case for the reported data
• adding a pre_sorted bool parameter to isotonic_regression with default True, which should allow for no change in performance for the primary class callers
• replaced the C reweighting approach with a data censoring pass s.t. all input values are within [y_min, y_max].

That said, as mentioned on the mailing list, I think we are still mixing use cases for the isotonic methods, and would recommend that we investigate docking and documenting against the R isotone package behavior:
#4185 (comment)