ignore: gracefully quit worker threads upon panic in ParallelVisitor

[cosmicexplorer]

Fixes #3009.

Problem

WalkParallel::visit() will nondeterministically hang if the ParallelVisitor::visit() implementation panics. This also occurs when providing a closure to WalkParallel::run(). Minimal repro is provided in #3009:

        WalkBuilder::new(path)
            .build_parallel()
            .run(|| Box::new(|_| panic!("oops!")));

The above code will nondeterministically hang, because of an infinite loop when no new work is available:

ripgrep/crates/ignore/src/walk.rs

Lines 1695 to 1707 in de4baa1

	loop {
	if let Some(v) = self.recv() {
	self.activate_worker();
	value = Some(v);
	break;
	}
	// Our stack isn't blocking. Instead of burning the
	// CPU waiting, we let the thread sleep for a bit. In
	// general, this tends to only occur once the search is
	// approaching termination.
	let dur = std::time::Duration::from_millis(1);
	std::thread::sleep(dur);
	}

Solution

1. Check the quit_now flag in our wait loop.
2. Catch any panic in the run() method and set quit_now before propagating the panic.

Breaking change: In order to ensure soundness, we also enforce that the filter method provided to WalkBuilder#filter_entry() is UnwindSafe. Users can circumvent this by wrapping in AssertUnwindSafe as needed.

Result

The added test panic_in_parallel() always succeeds instead of hanging.