Refactor AddDataMovePass (#822)

* Removed special case for permute >4D since this is fixed in tt-metal

* Typo fixes and conditional rearranging

* Rewrite _change_layout and _create_aligned_node for clarity

* Reduced set of operations used for _reset_to_default_layout
This passes lowering tests, but I am less confident in its generality

* Moved _call_to_torch_with_meta inside NodeAligner
Renamed variables and functions for consistency

* Added repeat to LAYOUT_CHANGE_OPS to get lowering tests to pass

* Removed _to_torch_with_meta
Reduced special case handling
Eliminated unneeded from_device calls

* Fixes an issue with tilized inputs to ttnn.split by converting to
row-major layout

* Removed function to check if ttnn.embedding is user of a given node
Updated TTNN_LAYOUT_CHANGE_OPS to separate TTNN_ROW_LAYOUT_OPS and
TTNN_HOST_ONLY_OPS

* Added to_layout call to convert max_pool2d output to tilized
Removed debug message that was cluttering logs

* Added to_layout calls for functions that might have row major layout from
ToTtPass

* One more fix for ones, ones_like, zeros, and zeros_like that default to
row major layout

* Added more special cases to get beit working, handle int inputs

* Added special case for GPT2 underflow issue

Author: jmalone-tt

torch_ttnn/backend.py

@@ -146,7 +146,7 @@ def aten_backend(
     passes = [
         ConstantFoldingPass(),
         ToTtPass(option.device, option.use_less_ttnn_op_types),
-        AddDataMovePass(),
+        AddDataMovePass(option.device),
         EliminateCoreopsPass(),
         CSEPass(),
         PermuteReshapeTuple(),