Reduce notebook memory footprint

[MichaReiser]

Summary

Reduce the memory footprint of NotebookIndex from O(lines) to O(cells).

The NotebookIndex stores a mapping from rows in the concatenated notebook document
to the relative row numbers within each cell and from absolute line number to the corresponding cell.

The way this is implemented today is by having a vector with one entry for every absolute row number
where the value is the index of the cell. While this allows O(1) lookup, it does require an entry for every single line.

This PR rewrites our representation (and uses one Vec instead of two) to only store the
start line number per cell and use a binary search to find to which cell a line number belongs.
This makes lookups slightly slower (from O(1) to O(log(n)) but reduces memory usage
from O(rows) to O(cells).

I noticed this improvement when working on ty's notebook support but decided to extract it into its own PR to make reviewing easier.

Test Plan

Existing tests

[github-actions]

ruff-ecosystem results

Linter (stable)

✅ ecosystem check detected no linter changes.

Linter (preview)

✅ ecosystem check detected no linter changes.

[dhruvmanila]

Suggested change

	/// This yields one entry per Python cell (skipping over Makrdown cell).
	/// This yields one entry per Python cell (skipping over Markdown cell).

[MichaReiser]

Thanks. I pushed the fix to #21175