Preserve weights through pickling and deepcopy

[vahid-ahmadi]

Partially addresses #300 (the pickle half — see the note at the end).

Problem

weights was a plain instance attribute, so it was not part of the pickle state:

pickle.loads(pickle.dumps(MicroSeries([1, 2], weights=[3, 4]))).sum()
# AttributeError: 'MicroSeries' object has no attribute 'weights'

That breaks to_pickle/read_pickle and anything shipping a weighted object across a process boundary — joblib, multiprocessing, Dask.

MicroDataFrame had a second, quieter failure underneath it. Its weighted aggregations are installed as per-instance closures by override_df_functions, which only runs in __init__ — a path unpickling skips. So even once weights were restored, an unpickled frame's sum was no longer overridden and fell through to pandas' unweighted implementation:

d = MicroDataFrame(pd.DataFrame({"x": [1, 2, 3]}), weights=[1, 2, 3])
pickle.loads(pickle.dumps(d)).sum()["x"]   # -> 6, expected 14 (no error)

Change

• _metadata = ["weights"] on both classes. pandas includes _metadata attributes in the pickle state, so weights round-trip.
• MicroDataFrame.__setstate__ reinstalls the aggregation overrides after unpickling.

Verification

Three new tests; all three fail on main (AttributeError for the Series, 6 != 14 for the frame) and pass with the change. Also verified by hand:

	before	after
pickle round-trip (MicroSeries)	AttributeError	14.0
pickle round-trip (MicroDataFrame sum)	6 (silently unweighted)	14.0
to_pickle/read_pickle	6	14.0
copy.deepcopy	14.0	14.0
groupby(...).sum() after unpickle	unweighted	14.0

Non-default indexes covered. make test — 65 passed on pandas 3.0.5 and pandas 2.3.3.

Scope

This is deliberately the low-risk half of #300. Declaring _constructor/_constructor_sliced — which is what would stop sort_values(ascending=False), fillna and sample from returning weightless plain pandas objects — is a bigger change: every pandas-internal construction would then run MicroDataFrame.__init__, including override_df_functions, so it needs a performance pass and a decision about how weights should be reindexed for row-changing operations. I've left #300 open for that, and noted there that the per-instance-closure design is what makes it awkward.