~10% memory regression in inference benchmarks due to 'ir: Fix incorrect renaming of phinode values (#52614)'

[Zentrik]

Feel free to close if not an issue/ determined to be acceptable etc.

The big regressions are mainly in the 'many_*' inference benchmarks so I imagine real world impact is a bit more limited. I bisected to #52614, from a brief look it's not obvious to me that this regression can't be avoided.
You can see the regressions here or in the report.