[WIP] Apply `&on-heap` selectively

[rsmmr]

In our generated C++ parsing code, so far we generally moved all units to the heap. This PR changes that to do so only for cases where we actually need it (primarily: if we wrap the unit into a reference somewhere, or if the unit type is cyclic). The assumption is that this should improve both performance and memory usage.

• Let optional's deref operator pass through "LHSideness".
• Prepare for moving parser temporaries to the stack.
• Add API to retrieve a type's dependent types.
• Rework emitting of C++-side type declarations.
• Apply &on-heap selectively only where needed.
• Change a set to an unordered_set.

[rsmmr]

Rebased on top of #1833, now using our new AST dependency tracking.

Test suite still fails because of #1831.

[rsmmr]

With #1833 and #1834 merged, this is now rebased on main.

Still not sure if we want this or not due to runtime performance concerns.