fix(core): release owning call-result/argument temporaries when a call throws

[Guikingone]

When a chained method call like $t->make()->boom() throws, the intermediate owning temporary from make() was leaked because longjmp bypassed the straight-line release code. This adds a runtime pending- cleanup stack (TLS global array) that records owning temporaries before a potentially-throwing call. If the call throws, __rt_throw_current drains the stack via __rt_eh_drain (calling __rt_decref_any on each entry) before longjmp. On the normal path, eh_pop removes the entry after the call returns.

New EIR ops: EhPush (1 operand) and EhPop (0 operands). New runtime helpers: __rt_eh_push, __rt_eh_pop, __rt_eh_drain (AArch64 + x86_64).
New globals: _eh_cleanup_stack (2048 bytes), _eh_cleanup_top (8 bytes).

Fixes #399

[nahime0]

Closing this draft. Do not close #399.

The leak is still real on main. This patch is not the fix: the cleanup stack is process-global (not TLS), would not survive fiber switch, already failed Linux PIE link, and the codegen paths it edits no longer exist.

Please do not rebase e61819bb. Salvage is the three destructor tests and the problem statement. A real fix belongs on current main, registered into the existing activation / fiber story, not a second global stack. #775 stays separate.