fix(distillfs): prevent duplicate watcher panic

[tianyuzhou95]

Summary

• Claim watcher ownership atomically so repeated starts cannot launch multiple watcher goroutines for one daemon lifecycle.
• Pass the current stop and completion channels into the watcher so cleanup cannot follow later mutations of the Daemon fields.
• Add a regression test that repeatedly starts one watcher and verifies a single clean shutdown.

Reproduction

On an unmodified main checkout at 5970730a, the regression test deterministically started multiple watchers and failed with panic: close of closed channel from distillfs.(*Daemon).watch.func1, matching the production stack.

With this change applied, the same test passed 100 consecutive runs and 100 consecutive runs under the race detector on the test node.

Verification

• go test -race ./pkg/imagemanager/distillfs -run '^TestDaemon_StartWatchOnlyOnce$' -count=100
• make check-fmt
• make vet
• make test

[hky1999]

LGTM — approving.

Verified beyond the diff itself: TestDaemon_StartWatchOnlyOnce passes -race -count=20 locally, and I also cherry-picked this commit onto a branch carrying #42 and ran the full Firecracker E2E suite there (including the daemon kill -9 recovery cases) — no regressions.

The fix is the minimal correct shape: CompareAndSwap makes watcher ownership atomic so repeated startWatch calls can never stack a second goroutine, and snapshotting stopChan/kickStop into the goroutine means the deferred cleanup closes exactly the channels this watcher was started with — later Daemon field mutations can no longer divert it onto a different channel, which is precisely the panic: close of closed channel production stack. The exiting watcher still resets watcherActive before the remount path can reacquire it, so the single-watcher invariant holds across restart.

CI is 13/13 green on c599e784b, the base is the current main head, and there is no overlap with #42. Ready to merge from our side — landing this before #42's rebase also removes one moving part from it.