[llvm] [AMDGPU] Set inst_pref_size to maximum (PR #126981)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

I had to change code to lower bound code size estimation to avoid prefetching of unmaped memory.

https://github.com/llvm/llvm-project/pull/126981