Refactor UringEngine: Decouple submission queue push from kernel submission

[hicder]

This change refactors the io_uring submission flow to improve thread safety and simplify the concurrency model by separating the submission queue (SQ) operations from kernel submission operations.

Changes:

• Renamed push_and_submit() to push_to_sq() to reflect that it only adds entries to the submission queue without triggering kernel submission
• Moved kernel submission (submitter.submit()) into wait_for_completion(), where it's called before cq.sync() to ensure pending submissions are sent
• Updated all call sites to use the new push_to_sq() method

Rationale:

• The previous architecture submitted I/O operations to the kernel immediately after pushing to the submission queue, which required careful locking semantics
• By deferring kernel submission until the completion queue is polled, we ensure that submissions are batched more efficiently
• This change continues the ongoing work to remove UnsafeCell usage and improve the thread safety of the io_uring engine
• The new flow is simpler: push to SQ → wait for completions → submit any pending operations to kernel → check completion queue

This is part of the series of commits to refactor UringEngine and eliminate UnsafeCell, following the pattern established in #614 and #615.