Improve the cost model for usage of in-memory range rowid / primary key filters

The current implementation of range rowid filters introduced in 10.4 (see MDEV-16188) used a quite simplistic cost model. In particular the cost of index only access was not taken into account properly. This task intends to remove all defects of the current implementation due to improper estimation of the gain promising by usage of rowid range filter. The full new cost model for usage of range rowid filters will be presented in the comments below.