Fix unsound assumption of chunks being disjoint

See https://github.com/ilan-gold/zarrs-python/pull/87#discussion_r1966169606

I assume we could still do things in parallel if we want.

We could create a DisjointViews (or so) struct that holds non-overlapping ArrayBytesFixedDisjointViews and has a safe fn add(&mut self, ...) -> bool that returns if a region was safely added (would not overlap with the already contained regions) or not. Then at runtime,

1. create a Vec<DisjointViews> by first trying to add each view to the first DisjointViews object, then the second, … until each view has been added.
2. go over the vec and handle each DisjointViews object in parallel.

There could be pathological cases where that doesn’t work well, but in practice, those shouldn’t happen

Start with the linked pull request discussion and trace the code using ArrayBytesFixedDisjointView and the assumption that chunks are disjoint. Determine how views are grouped before parallel handling; done means overlapping views cannot be processed together while non-overlapping work can still run in parallel.