Narrow index type made random access much slower

[robert3005]

We should investigate why #1558 made random access benchmark 50% slower

[gatesn]

Do you know specifically where the regression is? We do the widening with a cast in to_arrow, and I made the cast function fall back to canonicalize+cast, so we could just be missing a few cast implementations.

[robert3005]

I haven’t looked at profile. This is speculation given the changes merged that caused slow down

[gatesn]

So yeah, agree we should investigate and fix, not sure I agree with the proposed fix sans investigation