Unpin NumPy 2 and do necessary changes

[seberg]

Right now legate-boost pins to numpy<2 maybe just unnecessarily due to test failure.

The problem is a chain of things:

• legate-boost subclasses sklearn
• sklearns score method (unintentionally) uses partial Array API related logic (specifically it checks .device). Maybe only in an effort to give a nicer error message.
• On numpy < 2 those checks are just irrelevant, but on numpy >= 2 they trigger because cupynumeric don't have array.device == "cpu"

If we can create a more minimal reproducer, this is probably actually an skimage bug, because skimage shouldn't be sensitive to this type of thing without user opt-in.
On the one hand, that may be in the future, on the other hand such details are probably also still up to change in sklearn before removing the need for opt-in.

The simplest thing is to decide that just don't expect a legate estimator to work with NumPy inputs to .score() or convert to cupynumeric first to work around it (until sklearn may have changed their behavior).