Fixes a translation bug in translate_walk_restore() where it just skips GPR restores if it's in between application instructions. Since on AArchXX we have many of these cases of cross-app-instr spills (due to restore elision for back-to-back spills), the assumption that there will be no outstanding spills does not hold.
The solution is to move the condition to only encompass the stack pointer restore, while the GPR restores should always be done.
This was caught by the forthcoming common.broadfun-stress test, which will serve as the regression test once it is added after all other bugs it finds are addressed.
Issue: #4680 (closed)