PCC error for `mul` with sharded input

[yieldthought]

Painfully discovered and debugged during Grok-1 bringup. The mul operator (and inline * operator) do not support sharded inputs. They silently produce bad results for some of the cores. Oh, and their output is interleaved instead of respecting the input memory config and remaining sharded.

Added a unit test for this to /home/moconnor/main-metal/tests/ttnn/unit_tests/operations/test_mul.py (test_multiply_with_scalar_sharded) which fails on main. Available in branch yieldthought/mul_sharded_scalar.

[TT-BrianLiu]

The output memory_config is forced to interleaved always for scalar mul and it's related to this issue: https://github.com/tenstorrent/tt-metal/issues/7637

The default should be the memory_config of the input tensor, which is HEIGHT_SHARDED in this case. The following fix will get your test to pass and also have the output be sharded:

There are two issues:

• we need to find a way to add the above fix without breaking existing usage, since there are some models that expects the current behaviour.
• the interleaved output should also work, but it doesn't right now. I am looking into that

[TT-BrianLiu]

I identified what was causing the interleaved PCC issue. For sharded inputs, we need to be mindful of the shard orientation when distributing output work. What was happening was input shards are row-major and the work was being distributed col-major (which is the default).

This PR fixes that and also properly defaults output memory config to the input memory config when nothing is specified: https://github.com/tenstorrent/tt-metal/pull/9939

Side note: I picked up your test case and added it as well. I used ttnn.mul instead of * because I wanted to pass in the output memory config for testing.

[TT-BrianLiu]

Merged: https://github.com/tenstorrent/tt-metal/pull/9939