In atomic, if you exploded a rook, the castling rights were updated correctly but not the `unmovedRooks`, so if later one of your piece was taken on the same square (eg: white queen taken on H8), you would lose castling right on that side (White lost kingside castling after H8 capture).

The source of the bug is below, where the unmovedRooks from before were copied, which does not make sense? And in atomic resulted in an issue because the explosion code updated the `after` `Board`.
```diff
         val h2 = h1.copy(
           lastMove = Option(toUci),
-          unmovedRooks = before.unmovedRooks,
           halfMoveClock =
             if piece.is(Pawn) || captures || promotes then HalfMoveClock.initial
             else h1.halfMoveClock + 1
         )
```

The following change was unnecessary, but conceptually it makes more sense to have the unmovedRooks and castling rights copied from the original history. It should also prevent.
```
var castleRights: Castles      = h1.castles
var unmovedRooks: UnmovedRooks = h1.unmovedRooks
```

As the TODO say `Move.finalizeAfter` is very brittle, but more generally `Move` seems to contain far more information than it should, resulting in very weak typing because many functions are needed to "finalise", "finalise-after", etc, using many copies instead of types.

I put two tests, one specific but also the game in question until the unavailable castling to prevent larger regressions.

I have not run perft, will wait for CI.

close lichess-org/lila#14544