fix: propagate cancellation to parallel tactic subtasks via command-level snapshot tasks

This PR adds `Core.State.commandSnapshotTasks`, a separate array for snapshot
tasks that are reported at command level rather than tactic level. Unlike
regular `snapshotTasks` (which are consumed by tactic machinery and nested
into `TacticFinishedSnapshot.moreSnaps`), command-level snapshot tasks become
siblings in the snapshot tree, making their cancel tokens directly reachable
by `cancelRec` without waiting for the parent tactic to complete.

The `TacticM` parallel combinators (`par`, `par'`, `parFirst`, etc.) now
automatically register subtask cancel tokens as command-level snapshot tasks,
so parallel subtasks spawned during tactic elaboration are promptly cancelled
on re-elaboration. This removes the need for custom cancel-propagation
wrappers in `Try.lean`.

Also changes `asTask` to return `IO.CancelToken` instead of `BaseIO Unit`,
which allows direct registration of the token without a polling thread.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>