parallel scheduler bugfixes: shutdown, worker counts

[James-Gilbert-]

What

• Adds panic recovery for the developer-supplied do function
• Updates workersActive gauge when a worker exits
• Ensures that at least one worker is always started by NewScheduler()

Why

• Shutdown() relies on a fixed number of workers signaling via p.out. If fewer workers are available to signal, it can hang.
• More accurate metrics, prevents the silent loss of a worker if a task panics

Alternatives

• Immediately stop worker, clean up resources, and signal after recovering from a panic
• Use a sync.WaitGroup to track active workers, and manage lifecyle via context propagation
  • Maybe derive a taskCtx with timeout to mitigate head-of-line blocking in place of context.TODO()