Do not reuse global worker pool for files

[trowski]

This modifies the parallel driver to create its own internal pool by default, rather than reuse the global worker pool. A specific pool can still be passed to the constructor (though I assume this use is rather rare).

The AMP_WORKER constant is no longer used to prevent the blocking driver from being used. This way, a Task may still use async file access and get the behavior expected.

/cc @azjezz

[kelunik]

How do we prevent infinite ParallelFilesystemDriver instances now?

[trowski]

We don't, so tasks can use async file access. @azjezz mentioned this was an issue.

[kelunik]

But inside the worker we'll spawn another worker now and so on and never actually handle filesystem calls, no?

[trowski]

No no, FileTask which handles file system calls will use the blocking driver for the call. If the worker executes a user task, then another pool of workers might be spawned. I don't see infinite workers being likely (unless someone executes a Task within a Task, but that would be strange), but each worker having its own set of workers would be a possibility. Since by default we limit the number of workers to 8, the resource usage should still be within reason even on systems without eio or uv.