Fix BTree leaf split ownership transfer

[shuaimu]

Summary

• move BTree leaf-split medians out of MaybeUninit storage instead of bit-copying copyable owners
• consume removed BTree entries with MaybeUninit::assume_init instead of assume_init_read, so an Rc-like owner is transferred exactly once and the dead slot retains no hidden strong count
• encode both ownership fixes in the post-transpile patch pipeline and keep the patcher idempotent
• add destructor-count regressions for the first leaf split and middle/last removals
• load libc++ <vector> before module imports so the regression target builds without a forced include

Root cause

Both failures came from reading an initialized MaybeUninit<T> by copy where the BTree algorithm semantically moves T. Copyable owning types such as the generated Rc then left one untracked strong reference in dead node storage. The remove-path leak was also the cause of the previously observed ExtractIf use-after-free; it was not an independent defect.

Validation

• strict C++23 standalone suite: 478/478 passed
• full isolated ASan BTree suite: 478/478 passed with ASan errors fatal
• focused split and remove owner-drop accounting regressions passed
• post-transpile patcher is idempotent