SCX low bits are re-read per fetch, not once per line

[avivace]

Spun off from #626 to track an open discussion that shouldn't block merging that PR.

Context is:

@LIJI32 pointed out a subtlety about how SCX is sampled during background fetching:

You'd think the low bits of SCX would only be read once per line, because that's the logical way to do so, but unfortunately this isn't true. The fetched tile's X position is calculated (roughly) by adding SCX to LX and taking the 5 topmost bits, rather than adding the 5 topmost bits of SCX and LX.

We need:

• A test ROM (or other concrete evidence) demonstrating the per-fetch behavior.
• To fix the corresponding wording in Scrolling.md

cc @LIJI32 @ISSOtm