In SERVER-78354 we changed the implementation of algebra::transport so that it can handle trees of any depth without a stack overflow. However, this also made it slower, even for small trees.

Let's try making it faster. Some possibilities are:

• Change the memory layout of TreeCursor, or of the boost::vector intermediate results.
• Revert SERVER-78354 and solve the stack overflow (query_limits_test.js) a different way.

Note that this is not a CQF-only change, because SBE stagebuilders use algebra::transport.