fix(conn): handle overdue timers under Tokio budget exhaustion

[syszery]

This fixes a correctness issue where QUIC protocol timers could fail to fire within a scheduling round under load when Tokio’s cooperative task budget is exhausted.

drive_timer() previously relied on AsyncTimer::poll() to determine whether a deadline had elapsed. However, Sleep::poll() may return Poll::Pending for already-expired deadlines once the task has exhausted its cooperative budget. In that case, handle_timeout() would not be called until a later scheduling round, despite the deadline having already passed.

This addresses the timer starvation aspect of #753.

Fix

The timer logic is changed to:

• Always check runtime.now() >= deadline before polling the async timer
• Use the async timer only for future wakeups (waker registration)

[djc]

(Not sure why this is draft? Will hold off on reviewing for now.)

[syszery]

Thanks. I just double checked something. The PR is ready for review now.

[djc]

I'd like to have this split in two commits:

• One to use let-else instead of match
• Another to change the logic

A bunch of comments were removed -- were those removals motivated or just LLM collateral?

[syszery]

The PR is split into two commits now.

The comment changes were intentional:

• The first two described the old tangled match structure, which I think is easier to follow after the let...else extraction.
• The "store expiration time" comment was just restating the line below it.
• The final two became slightly inaccurate after the fix.