[llvm] [ARM] Only change mask if demanded bits says we can optimize (PR #165106)

[via llvm-commits]

AZero13 wrote:

This is the best I can do without regressions @davemgreen 

To remove the workarounds I have to improve the ubfx code which involves porting getUsefulBits from AArch64, which is a very large effort. 

https://github.com/llvm/llvm-project/pull/165106