rsx: Vertex program analyser improvements

[kd-11]

• Fixes a bug when the program entry is at the end of an ubershader block. In this scenario, the entry is the dispatch block and the subroutines sit above it which could cause a crash if there is padding in the structure.
• More aggressive DCE. In case a loop is detected, we don't need to walk the loop again as it will always end the same way. Previously, we walked over the blocks and consumed the tail sections which could add hundreds of instructions of unreachable code at the end of ubershaders.

Closes #13443