BUG: .transform(...) with "first" and "last" fail when axis=1

[rhshadrach]

Part of #45986 (first/last)

• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

This removes warnings of dropping nuisance columns when using e.g. .transform("mean"). This makes it consistent with .agg, and in #46072 both transform and agg will warn about the default switching from numeric_only=False to numeric_only=True. Doing this first will make #46072 slightly easier.

Not falling back to _transform_item_by_item is also better for performance. For the benchmark, I disabled the warning being emitted in main to make sure that wasn't messing up the results.

size = 1_000_000
df = pd.DataFrame(
    {
        "A": size * ["foo", "bar"],
        "B": "one",
        "C": np.random.randn(2*size),
        "D": np.random.randn(2*size),
    }
)
%timeit df.groupby("A").transform("mean")

# This PR
108 ms ± 966 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
# main, but with size set to 100_000 (1mm was taking very very long)
920 ms ± 7.95 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[rhshadrach]

In order to still hit the warning this was testing for, I needed to switch from a duplicated float column to a duplicated string column, as we no longer warn in the case of the duplicated float column. But using a string column would always fail with SeriesGroupBy, which is why duplicates=False was removed.

[jreback]

thanks @rhshadrach