sorted_set: union defensively copies both trees and recounts the result size

[bobzhang]

sorted_set/set.mbt SortedSet::union (the TODO: optimize this. Avoid counting the size of the set at ~line 175):

1. copy_tree is called on both operands up front. The copies are required by the current mutable design (add_node/delete_node/rotations mutate nodes in place, so subtree sharing would corrupt the source sets) — but it means union is O(n+m) allocations before any merging starts.
2. After the split/join merge, the result size is recomputed by a full each traversal because nodes store only height, not subtree size.

Possible directions:

• Store subtree size in Node (as the immutable immut/sorted_set does) — kills the recount and gives O(1) length for subtrees; costs one word per node and bookkeeping in every rotation.
• Count duplicates during split (a duplicate is discovered exactly when split hits comp == 0), then size = size1 + size2 - dups — no node layout change, but split needs a counting variant.
• For the copy overhead: iterate the smaller set and add into a copy of the larger — O(n + m log n) with one copy instead of two and no recount (add maintains size). Loses the asymptotic edge of split/join union for balanced sizes.

Related: symmetric_difference had the same shape and is addressed by #3819 with a merge-walk; intersection (each + contains) may also benefit from a merge-walk.