Fix to_prev_sibling_byte losing the child of a list node location

[imlvts]

A location with both a value and a child is two same-key slots in a LineListNode. The previous-sibling lookup only checked the first matched slot for the child, so the zipper landed with child_count 0. Check both.

[imlvts]

https://github.com/Adam-Vandervorst/PathMap/actions/runs/35158141349#summary-105008105822

Again, +57.2% superdense_drop_bench/200, but overall result not changed?

[luketpeterson]

@imlvts I promoted your test to all zippers and added another test. That turned up an ACTZipper bug.

I committed a fix for that bug, but I wanted your review in case the fix isn't agreeable. I factored it the way I did because having the focus at a non-existent path is an edge case in practice and I didn't want the case to slow down anything that is doing iteration over existing items.

[luketpeterson]

Again, +57.2% superdense_drop_bench/200, but overall result not changed?

I feel like the drop bench results are so noisy, I'm of a mind to remove those benchmarks altogether.

If I thought this was the final shape of the TrieNode contract, I'd optimize the function more, but I don't think it's worth any time right now because there are a number of changes I want to make.

We don't really have any benchmarks that exercise backwards iteration, but maybe we should.

[imlvts]

This conflicts with #113, and I will update that later. The way you did it is better.
I'll look into benchmarks, my understanding is that it's only act value count that behaves weird.
If you're happy, we can merge this.

[luketpeterson]

Opened #120 to track the benchmark discussion