PERF: faster indexing for non-fastpath groupby ops

[jbrockmendel]

Per discussions about removing libreduction code, this is part of an effort to make the non-libreduction path more performant.

Performance comparisons are done by disabling fast_apply entirely and taking the two most-affected asvs:

import numpy as np
from pandas import DataFrame

N = 10 ** 4
labels = np.random.randint(0, 2000, size=N)
labels2 = np.random.randint(0, 3, size=N)
df = DataFrame(
    {
        "key": labels,
        "key2": labels2,
        "value1": np.random.randn(N),
        "value2": ["foo", "bar", "baz", "qux"] * (N // 4),
    }
)

%prun -s cumtime df.groupby(["key", "key2"]).apply(lambda x: 1)
PR -> 0.263 s
No optimization -> 0.308 s
master -> .039 s

%prun -s cumtime df.groupby("key").apply(lambda x: 1)
PR -> 0.083 s
No optimization -> 0.127 s
master -> .012 s