Validate max_num_features is positive in plot_importance

[Nityahapani]

Validate max_num_features is positive in plot_importance

Summary

plot_importance() implements the "top N features" selection via Python
negative-index slicing:

tuples = sorted(tuples, key=lambda _x: _x[1])[-max_num_features:]

This has two undocumented, confusing edge cases:

• max_num_features=0: Python's [-0:] is the same as [0:], i.e. the
  entire list — so 0 silently means "show all features", the opposite
  of what a reasonable reading of "maximum number of top features
  displayed" (the parameter's own docstring) would suggest.
• max_num_features<0: [-(-N):] becomes [N:], which skips the
  first N lowest-importance entries after sorting ascending, i.e. it
  shows the least important features rather than raising an error or
  showing the top ones. E.g. max_num_features=-3 on 5 features shows the
  2 least important ones.

Neither case has any sensible interpretation given the parameter's
documented meaning ("Maximum number of top features displayed"), so
there's no reasonable value for a user to fall back on — an explicit error
is the right behavior.

Change

     tuples = [(k, importance[k]) for k in importance]
     if max_num_features is not None:
-        # pylint: disable=invalid-unary-operand-type
+        if max_num_features <= 0:
+            raise ValueError(
+                "`max_num_features` must be a positive integer, got "
+                f"{max_num_features}. Note that due to how the top-N "
+                "selection is implemented, 0 or a negative value does not "
+                "mean 'show no features' -- it previously produced confusing, "
+                "undocumented results (0 silently showed all features; "
+                "negative values showed the *least* important features "
+                "instead of the most important ones)."
+            )
         tuples = sorted(tuples, key=lambda _x: _x[1])[-max_num_features:]
     else:
         tuples = sorted(tuples, key=lambda _x: _x[1])

This also removes the existing # pylint: disable=invalid-unary-operand-type
comment, which was suppressing pylint's (correct) inability to prove
max_num_features was numeric before negating it — the new explicit check
makes that provably safe, and I confirmed pylint no longer flags it once
the suppression comment is removed.

Verification

• None, and positive ints (1, 5, 100): unchanged, valid
• 0, -1, -3: now raise ValueError with a clear message instead of
  silently producing wrong results

mypy, ruff, and pylint (specifically re-checked for
invalid-unary-operand-type) all pass on the modified file.