[PERF/ENH] Index.intersection does more hashing work than necessary

[wence-]

Index intersection performs an inner merge of the unique values of the left and right indices (the unique is done so that indices with repeated values don't blow up the memory footprint). This does a full hash of both indices, then the merge (hashing again). Finally, if requested, the result is sorted.

This could be replaced, I think with positive performance effect by either:

• leftsemi join + drop_duplicates
• libcudf.search.contains + apply_boolean_mask + drop_duplicates

One would have to think through the consequences of either of these wrt any ordering guarantees we might want when sort=False (possibly gated behind pandas-compat mode).

This applies mutatis mutandis to MultiIndex.intersection too.