This adds a new strategy for paced mutations that enables parallel execution
of mutations that affect different records, while still serializing mutations
that share dependencies (based on globalKey).

Key features:
- Mutations on different records run in parallel (no artificial serialization)
- Mutations on the same record are serialized in order
- Supports optional custom dependencies via getDependencies callback
- Rollback cascade still works correctly across dependent mutations

This addresses the use case where users want maximum parallelism without
sacrificing correctness - e.g., editing multiple independent items shouldn't
block on each other, but edits to the same item must be ordered.

Example usage:
```ts
const mutate = usePacedMutations({
  onMutate: (vars) => collection.update(vars.id, ...),
  mutationFn: async ({ transaction }) => api.save(transaction.mutations),
  strategy: dependencyQueueStrategy()
})

// These run in parallel (different IDs):
mutate({ id: 1, title: 'A' })
mutate({ id: 2, title: 'B' })

// This waits for the first one (same ID):
mutate({ id: 1, title: 'C' })
```