refactor: remove cancelled counter via O(1) abort removal

[Kikobeats]

The waiter queue was singly-linked with lazy deletion: an aborted waiter could not be unlinked in O(1), so it stayed in the queue as a tombstone and a cancelled counter tracked how many were pending. release() walked past them and awaiting() subtracted them. When aborts outpaced releases, both the queue and cancelled grew without bound.

Make LinkedList doubly-linked (prev pointer + O(1) remove(node)) so abort unlinks its own node immediately. This deletes the cancelled counter, the waiter.cancelled flag, the skip-loop in release(), and the subtraction in awaiting(). The queue now holds only live waiters, so awaiting() === size().

Public API unchanged; all 17 tests pass at 100% coverage.

---

Note

Medium Risk
Changes concurrency wait-queue behavior (abort, release, awaiting) though the public API and tests are unchanged; incorrect unlinking could affect slot handoff order or counts.

Overview
Lock queue internals are reworked so aborted waiters are removed from the queue immediately instead of staying as tombstones.

The internal LinkedList is now doubly-linked (prev + remove(node)), enqueue returns the node, and AbortSignal handlers call queue.remove(node). That drops the cancelled counter, per-waiter cancelled flags, the release() loop that skipped dead entries, and the awaiting() adjustment—awaiting() now matches live queue length only. release() only dequeues the next live waiter.

CI sets fail-on-error: false on the Coveralls step so coverage upload failures do not fail the workflow. A cancel test comment is updated to match the new behavior.

^{Reviewed by Cursor Bugbot for commit f47ca74. Bugbot is set up for automated code reviews on this repo. Configure here.}