[single] Refactor invoke_thread to reduce goto statements

[songgot]

Refactor invoke_thread to remove goto statements and simplify control flow

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR removes the wait_for_next label from invoke_thread, renames exit to exit_thread, hoists single_h->status = status above the error/JOIN check, and turns the success path into an else branch. I traced every path in old vs new: the mutex is held on entry to every goto exit_thread, after the re-lock following __invoke, and on natural loop exit, and exit_thread unlocks exactly once. The g_cond_wait loop, invoking flag, output freeing on error/join vs __process_output on success, the alloc_output-guarded cleanup, broadcasts, and NULL in/out handling are unchanged. Merges cleanly onto current main. Existing tests (invoke_timeout, parallel_runs, close_while_running) cover all three thread paths. Minor comments only.

1. [Low] Small semantic change not mentioned — c/src/ml-api-inference-single.c:545: single_h->status = status is now also stored when state == JOIN_REQUESTED after __invoke. Harmless (a still-waiting _ml_single_invoke_internal gets the real result instead of a stale one), but the PR is described as a pure refactor. Either keep the assignment where it was (just before the if (single_h->state == RUNNING)) or mention the change in the commit message.
2. [Low] Inaccurate comment — :554 /* If join requested, exit immediately without broadcast */: exit_thread does broadcast in the JOIN_REQUESTED branch, and that broadcast is required to wake ml_single_close waiting on invoking. Reword to e.g. /* Join requested: release data and broadcast at exit_thread */.
3. [Low] Comment style / lost invariant — :562 /*Reset state and notify */ is missing the space after /*; and the original /* Do not set IDLE if JOIN_REQUESTED */ at the exit label was replaced by /* Cleanup resources on exit */, dropping a non-obvious invariant that ml_single_close / invoke_internal rely on. Please restore it alongside the new comment.
4. [Low] Stale Android CI failure — the PR's Android Build Test (2025-12-26) failed on all four ABIs; logs have expired and the diff is platform-independent control flow, so this is almost certainly transient. Please rebase onto current main and re-run CI before merging.

No back-door or suspicious behavior found.