Fix race on cancelled incremental queries

[emcfarlane]

This fixes a race found with repeated cancellations in the LSP from testing. Auditing the surround found a couple of issues around pending tasks:

• Results computed under a cancelled context were cached permanently
• A panicking leader only cancelled its own Run, blocking concurrent Runs
• A leader cancelled while queued on the semaphore left its key pending, blocking concurrent Runs
• Diagnostics from a failed attempt were reported again on retry.

The fix marks results as completed (done, closed) and withdrawn (unpublished, closed). Withdrawal wakes waiters, which retry and elect a leader.

Cycle detection now walks the waiters list, instead of the deps graph. This avoids misclassification from a concurrent Run where a previous one was a cycle. Edges carry the requesting query, so cycle errors correctly identify the trail.

Also: WithTimings was never populated (recorded nothing); Run's diagnostic collection skips uncompleted tasks; FDS no longer mutates the memoized Link slice; transferFrom tolerates holds lost to cancellation; the reentrant-Run guard no longer mutates a shared slice.