Lowering aten.native_group_norm.default

[swimdi]

Ticket

#538

Problem description

Support the lowering of aten.native_group_norm.default

I implement it accroding to test_group_norm.py on tt-metal

Originally I reference the test_group_norm_with_block_sharded_v2_8x8_grid_tile_layout part, but some cases accuracy failed

"input_shape, num_groups",
[(1, 320, 32, 32), 32]
[(2, 320, 64, 64), 32]

And after I set the input_tensor layout as TtnnRowMajorLayout, its accuracy pass

And there still some case failed, I issued at #555

What's changed

• Support lowering aten.native_group_norm.default
• Add aten_native_group_norm_default_blocklist
• Add test_group_norm.py

btw, only stable diffusion v2 & retinanet_resnet50_fpn_v2 use group_norm, so maybe this op's priority can lower

I'll merge #532 first and then this PR

[ayerofieiev-tt]

And after I set the input_tensor layout as TtnnRowMajorLayout, its accuracy pass

Please fire a ticket. There should be no requirement to change from Tile to RM for PCC to pass. Sounds like a bug to me.

[swimdi]

And after I set the input_tensor layout as TtnnRowMajorLayout, its accuracy pass

Please fire a ticket. There should be no requirement to change from Tile to RM for PCC to pass. Sounds like a bug to me.

that case is mentioned in #555 , or you prefer an individual issue@@?

[jerrysky3]

LGTM on the group norm parts added by this PR

[jerrysky3]

Maybe add a TODO to remind us the shard shape computation can be improved later