Investigate low accuracy in mixer_b16_224.goog_in21k (42.74)

[amalbasaTT]

Accuracy at the moment: 42.74
https://github.com/tenstorrent/pytorch2.0_ttnn/blob/main/docs/models/mixer_b16_224.goog_in21k/input_variations.md

[amalbasaTT]

ttnn.layer_norm causes low PCC is mixer_b16_224.goog_in21k model. Replacing it with it's fallback varient raises PCC to 0.9276377837318575. New accuracy can be reproduced by following next steps:

1. checkout branch amalbasaTT/mixer-test-file
2. Run command:

pytest tests/models/timm/test_timm_image_classification.py --gen_op_accuracy_tests

3. Run command:

python tests/autogen_accuracy_tests/mixer_b16_224.goog_in21k_code_ln.py

4. Output will contain PCC errors for every faulty ttnn.layer_norm along with model accuracy when all of them are converted to their fallback version.

[amalbasaTT]

Reducing ttnn.gelu and ttnn.layer_norm to fallback increases PCC from 0.92 to 0.94 (tenstorrent/tt-metal#20562)