Include location bias into rescoring

[lonvia]

This extends the rescoring algorithm from #1041 to also take into account the score from the location bias. Thus the final OpenSearch score is now neatly split into match score, importance score and location bias score. The location bias score is added such that if nothing suitable is found in the focus area, the rescoring behaves exactly as if no location bias was set at all.

This also changes the size of the focus area (larger for large zooms, smaller for smaller ones) and the default zoom (14 -> 12) and location_bias_score (0.2 -> 0.4).

The PR also adds the output of OpenSearch's explain to the debug output for search. That can be useful to understand how the score came to be.

@henrik242 would you mind checking the effects on your data?

Closes #1024.

[henrik242]

I'm seeing a couple of unexpected results. Searching for Schw 10 with a focus in Oslo previously gave me Schweigaards gate 10, Oslo (distance: 1 km) as the first result. Now I get Schwartz gate 10, Drammen (distance: 36 km).

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit 5949947. Configure here.}

[lonvia]

Added a couple of adjustments to balance out the various use-cases:

• Added a minor decay in the area of focus, so that results closer to the center come first when all other things are equal.
• No longer make rematch score depend on length of place name. If doesn't make sense to give "Neuhausen" a lower score than "Neuss" when searching for "Neu" just because the former has more letters.
• Don't use the strong exact match rescoring when the query contains a comma. A comma indicates that the query is made up of multiple terms and thus shouldn't match against the exact name.
• When rescoring ends in a 0 score, still give it a small score relative to the place's importance. That allows for a bit of typo-tolerant matching for well-known places.

I've also adjusted the focus radius computation some more. The result seems to be fairly decent for the planet test set. We might do something with getting more result candidates and doing some post-filtering in order to get result list that don't have obvious outliers down the list but that is something for the next version.