Use view types for sort-heavy workflows

When we are going to do a sort (training an index, user requested order-by, etc.) we should be using view types (binary view and string view). These sort much faster and, perhaps more importantly, they don't suffer from the same overhead issues (and sort node OOMs) that the other variants do.

In fact, it's probably safer to err on the side of using view types too much. The only time the view types are a penalty is when we have small strings and some kind of text-processing kernel (e.g. string contains).

We will also likely need to convert back to dense strings for the user if they request it.

This is related to #7069

Start by tracing the sort-heavy workflows described in the issue, including index training and user-requested order-by operations, and review related issue #7069. The work is complete when these paths use binary or string view types where appropriate, while user-facing dense-string results are preserved when requested.