Added a helper class _PlacementCache that maintains all the indices needed for fast near-constant-time op placement during append, speeding up a worst-case quadratic perf behavior when constructing circuits via repeated append calls. Any non-append modification of the circuit simply nulls out the placer, and subsequent modifications just fall back to the (existing, slow) placement index search from then on.

Updated transformer_primitives.apply_map_func to also use this class, since much of the code was duplicated.

There's probably room to support additional insert strategies, or to allow for manual recreation of a nulled-out circuit placer, etc., to increase the applicability of this. But for now, append is the most common way to create circuits, and this PR handles that case.

As mentioned in the test, this change fixes the placement perf, which is most egregious for long narrow circuits. (For short, wide circuits, the cost of recreating the (immutable) Moment for each new op added to it dominates, which this PR does not address.)

Also note that this change has no effect on performance of constructing circuits by passing in a list of ops into the Circuit constructor, which is already fast for both long and wide circuits. The only reason this PR is needed is because many of our tutorials demonstrate using repeated append calls for constructing circuits rather than passing the ops into the Circuit constructor, and it appears that many users do the same.