Don't leak __tmp_weights column out of groupby

[MaxGhenis]

Summary

MicroDataFrame.groupby used to do self["__tmp_weights"] = self.weights and never clean it up, so after a single df.groupby(...) call the caller's DataFrame permanently carried the weight column. Any later df.sum() or iteration over columns then picked it up as data.

Fix: stage the weights onto a copy before calling super().groupby() rather than mutating self. All existing groupby plumbing (_weights_groupby, per-column MicroSeriesGroupBy) keeps working.

Reproduction

import microdf as mdf

df = mdf.MicroDataFrame({"g": ["a", "a", "b"], "v": [1, 2, 3]}, weights=[1, 2, 3])
_ = df.groupby("g").sum()
list(df.columns)   # was ['g', 'v', '__tmp_weights'], now ['g', 'v']

Test plan

• Existing 51 tests still pass
• New test_groupby_does_not_leak_tmp_weights_column covers single-column and list-of-columns groupbys and verifies the weighted aggregation result stays correct.

[MaxGhenis]

Self-review. Verified:

• Old path self["__tmp_weights"] = self.weights mutated caller; new path copies pd.DataFrame(self).copy() into staged and adds the weight column there. self.columns is untouched.
• gb = staged.groupby(...); subsequent gb["__tmp_weights"] still resolves because the column lives on staged (which the GroupBy object holds).
• Loop for col in staged.columns iterates over self.columns + __tmp_weights; per-column MicroSeriesGroupBy rewrapping is unchanged.
• CoW note (pandas 3): pd.DataFrame(self).copy() returns a detached copy, so even under CoW the staging column mutation doesn't reach self. Safe.
• Regression test verifies single-column and list-column groupby don't leak __tmp_weights, and that the weighted sum result is still correct.
• CI green Py 3.9-3.13.