Avoid DataFrame conversion in MultiIndex.from_pandas

[mroeschke]

Description

This uncovered a bug where MultiIndex.levels could be a list of Series instead of a list of Index.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wence-]

On line 175 we already do partial bounds-checking. The treatment of -1 is AFAICT undocumented pandas behaviour. That is codes do not do normal indexing but instead -1 is for "missing value".

But let us move all of the validation into one place:

import cudf._lib as libcudf
from cudf._lib.types import size_type_dtype
import numpy as np

lo, hi = libcudf.reduce.minmax(code)
if lo.value < -1 or hi.value > len(level) - 1:
    raise ValueError(...)
if lo.value == -1:
    # Now we can gather and insert null automatically
    code[code == -1] = np.iinfo(size_type_dtype).min
result_col, = libcudf.copying.gather([level], code, nullify=True)

[mroeschke]

Nice find. Was able to consolidate the codes validate in this loop

[wence-]

FWIW, I am not sure that codes can be np.int64 it should probably be size_type_dtype. But let us not fix that here.

[wence-]

Minor suggestion to retain the pandas issue reference in the xfail, otherwise looks great, thanks!

[galipremsagar]

@mroeschke can we drop the REF: prefix from the PR title? A while ago we decided not to have prefixes in the PR titles as they will be added to the commit history.

[mroeschke]

/merge