pack_untilize_block causes PCC error on BH

[skrsticTT]

Downsample op does not work on BH because of PCC errors.

It turns out that in compute kernel ttnn/cpp/ttnn/operations/pool/downsample/device/kernels/downsample_fast_pack_untilize_compute_kernel.cpp method pack_untilize_block (second call, first and third are not used in my repro) could work improperly. On WH, it just shuffle datums in a tile, every datum stays in a tile. But on BH, it looks like some datums are lost. Also it looks like first 2 faces are OK, and last 2 faces are not OK.

pack_untilize_block is also used in kv_cache op, which is used by a few models (these models are still not tested on BH).

In repro, I printed tile before and after that call to see changes. Also datums are set to be 0,1,2,3... so it is easy to spot a difference.
Repro branch: skrstic/downsample-repro
Repro command: TT_METAL_DPRINT_CORES="(0,0)" pytest tests/ttnn/unit_tests/operations/test_downsample.py