Exchange columns more efficiently

[frankmcsherry]

The columnar trie datastructures used in examples/columnar.rs are grouped by key, and also need to be exchanged by (a hash of) the key. We can do this without departing from the columnar representation, which this PR introduces. It also inlines a hot function in ord_neu.rs.

[frankmcsherry]

Qualitatively, performance ticks up on larger batching sizes where we spend relatively more time exchanging, but is harder to notice for small batch sizes when we spend most of our time maintaining arrangements (which this does not affect).

On master:

     Running `target/release/examples/columnar 10000000 1000000 -w4`
5.86842775s     loading complete
5.8684315s      loading complete
5.8684335s      loading complete
5.8684325s      loading complete
15.462424042s   queries complete
15.462429375s   queries complete
15.462428875s   queries complete
15.462431208s   queries complete
15.464183083s   shut down

vs this branch:

     Running `target/release/examples/columnar 10000000 1000000 -w4`
5.175452958s    loading complete
5.17545875s     loading complete
5.175458208s    loading complete
5.17545875s     loading complete
13.752099208s   queries complete
13.752102208s   queries complete
13.75210275s    queries complete
13.752103208s   queries complete
13.762354208s   shut down