Make PDSH benchmarks (or most of them) work

[seberg]

Some already work, some of the fixes are work in progress, but to organize the missing pieces a bit, creating an overview issue:

• q1,q21: missing groupby-agg length (others work). q21: now fails due to slice not working on string arrays.
• q2, q9, q20: require (some) string functions.
• q4, q18 requires "semi" join. (We can already map that to a left join probably, but I suspect we should just implement the semi join.
  • (should add a test for duplicate handling on the RHS, I am not quite sure how it is defined for a semi-join.)
• q7: Requires union (not immediately sure if that can be mapped to a join).
• q8, q14: Needs libcudf copy_if_else and enable ternary operation in polars.
• q11, q22: uses a conditional join, have to see if that can be done in the polars layer or needs deeper support.
• q12: (not sure currently a small bug, but something else may crop up)
• q13: "boolean not" failing, likely a small issue.
• q15, q17: "round" missing.
• q15: running into a broadcast issue now (may or may not be small one).
• q19: isin missing.
• q2: missing mask_nans should be easy to add.
• q10, q21: missing string_array.slice() for the final .head(), only. Likely being fixed in legate/legion.
• q20: missing unique (may be able to hack it via join in a pickle)

Still a lot not working (to be honest, more than I thought), but many should not need much.
(Overall, there is some slowness going on, but not sure where yet.)