fix(tiling): fall back to a smaller margin instead of dropping tiles

[EliHei2]

When the inner-tile negative buffer uses too large a margin, small tiles shrink to empty. Previously those tiles were removed from the spatial query, silently dropping every geometry they contained (a frequent cause of large numbers of unassigned transcripts on dense/small-tile datasets).

Instead, progressively halve the margin (down to 0) until every tile survives the buffer, emitting a warning with the reduced value. No behavior change when the requested margin already leaves all tiles non-empty.

What to review:

• src/segger/data/tiling.py: the buffer/fallback loop in the query method.
• The loop is bounded (margin strictly halves; terminates at 0 = unbuffered tiles, which can never be empty), so it cannot spin.
• Trade-off: a reduced margin means slightly less inner/outer separation for the affected tiles, which is strictly better than dropping their geometries.

[Tobiaspk]

I prefer this over the current approach to dropping empty tiles entirely. I'm good to move forward with this. But going forward, let's update this with a more dynamic approach that reduces empty tiles in the first place.

@andrewmoorman suggested transcript-density-derived margins. Please also add your comments on this so we're on the same page.

[EliHei2]

I like transcript-density-derived margins, but that's another dev, for now for not entirely missing the tile could we put this in, then add that as a future TODO?

[Tobiaspk]

Looks good. Let's make sure to update this in a future revision with a density based estimator.