Diff D17640 introduced a subtle and undetected bug: It ran the tx size estimation within tests using real signatures instead of EccDummy. However, the "leftover" calculation always uses EccDummy.
Since Schnorr signatures are fixed length, this was undetected. But for ECDSA signatures, the tests fail, as "real" signatures can have a different length than dummy signatures used for leftover calculation.
This diff restores the old behavoir, and additionally tests size estimation using ECDSA signatures to ensure we cover behavoir this in our tests.