[FEA] Check dtype requirements on multiindex codes

[wence-]

Is your feature request related to a problem? Please describe.

(Seen as part of a review of #14470).

Multiindex codes and levels are effectively a categorical encoding of the columns of the multiindex entries. The codes are used to index the levels. As such, they should probably have type equivalent to cudf::size_type. Currently, however, they are a int64. This is a larger memory footprint than necessary. Moreover, it (in some constructor circumstances) necessitates more copies than necessary.

Describe the solution you'd like

Use correct dtype. Since the public codes and levels properties wrap the results in pandas FrozenList objects to mimic the pandas API, it may be possible to just store the codes/levels pairs as CategoricalColumns internally, rather than the current structure.

Describe alternatives you've considered

n/a

Additional context

n/a