Partitioner to not delegate mixed dtype operators

[mcr229]

Problem

Currently there is no check that the XnnpackPartitioner does not partition mixed dtype ops for example a multiply that takes in a bf16 and fp32 input (these are not supported by xnnpack).

Context

In the XnnpackPartitioner we have configs for operators that describe how operators should be partitioned, and what constraints they must pass. The below link should show constraints that must be passed by all operators

executorch/backends/xnnpack/partition/config/xnnpack_config.py

Line 116 in 1a9a59b

def check_common_constraints(

We should add a check in this constraint to fail partitions that take in mixed dtypes as inputs. The refactoring done in this issue might be a good way to help enforce this more easily for binary inputs. #9024

Verification/Testing

Add a test with a pytorch module for operations that take in different dtypes (fp32, bf16, fp16, int8). Check that these are not partitioned:

Inverse of this:

executorch/backends/arm/test/ops/test_cat.py

Lines 69 to 70 in 1a9a59b

	.check_not(["executorch_exir_dialects_edge__ops_aten_cat_default"])
	.check_count({"torch.ops.higher_order.executorch_call_delegate": 1})

in that we want to check that there are no call_delegate nodes and the original node is still in the graph

Test for the following representative ops:

• cat
• mul
• add
• div
• sub

Resources

https://discord.gg/a9r5KZDNfZ

cc @digantdesai @cbilgin