Multichip models can't use rms_norm op that we currently support because they reduce over multiple chips.

We need to add fused_rms_minimal op to TTIR and TTNN dialects to be used in these cases.

Note: needs this commit on tt-metal from 10.01.2026. to be uplifted in tt-mlir.