Investigate why reduction fusions are not being sent to rocMLIR

DOR (Definition of Ready)

This can be worked on now.

Description

As a software developer I want to investigate why reduction fusions don't seem to be going to rocMLIR. Here is an example from running unet with the following command:

MIGRAPHX_ENABLE_MLIR_REDUCE_FUSION=1 MIGRAPHX_ENABLE_MLIR_GEG_FUSION=1 MIGRAPHX_MLIR_ENABLE_SPLITK=1 MIGRAPHX_MLIR_USE_SPECIFIC_OPS="fused,attention,dot,fused,convolution" ./AMDMIGraphX/build/bin/driver perf /mnt/nas_share/migraphx/models/sd/stable-diffusion-2-onnx/unet/model.onnx --fp16 --input-dim @sample 2 4 64 64 @timestep 1 @encoder_hidden_states 2 64 1024

In looking at the output from this run, I see lots of kernels like: gpu::code_object::mul_reduce_sum_kernel: 0.138859ms / 8 = 0.0173574ms, 1%} that don't get sent to MLIR.

DOD (Definition of Done)

• Understand why reductions are not being sent to rocMLIR
• Implement a fix (if necessary)

Run the supplied driver perf command and inspect whether gpu::code_object::mul_reduce_sum_kernel entries are routed to rocMLIR. Trace the reduction-fusion decision from that output; done means explaining why these reductions are excluded and implementing a fix if necessary, then confirming the behavior with the same workload.