Now that I'm creating the inputs dynamically, it doesn't help to have a class for labeled inputs.
And `self._inputs_by_letter[component_letter] = labeled_inputs[-1].query_one(Input)` failed, because it's not mounted yet.
It didn't actually cause an error at runtime because
format_from_extension just uses os.path.splitext, which uses os.fspath
which converts Path objects to path strings.
I had kept these around because several of these I'll want to use,
and the auto-import feature wasn't working, but I've since found a fix
for that (in 453111269e).
I'm also adding rich here to fix auto-imports from rich.
This brings it down from around 20 seconds to less than a second,
for moderately sized documents (i.e. default size).
For larger documents, it can still be slow.
It's still O(n^2) complexity, but more specifically, it's
(n^2 - n) / 2 iterations instead of n^2.
I don't know much about the performance of `enumerate()` or slicing.
It handles the pathological character grid test case SVG I devised,
as well as loading the SVG as saved by the app.
It's a little weird that I managed to get the pathological case working
before the simple rigid grid, but it was what I was shooting for,
and the main problem with the rigid grid was the spanned rects,
which I didn't think of when making the test case SVG.
It does have one problem left; characters that are wider than a cell,
when saved as SVG, take up two cells, but in my SVG loading code,
they're just placed in one cell, so, when saving and loading,
it shifts content to the right of the wide character rightwards,
expanding the region of the background from the wide character's cell.
But it converges now.
