Improve simplify_algebra

Problem Description

Once the mul_add_kernel, add_kernel and add_kernel are horizontally fused with the mlir_dot_add and mlir_slice_mul_reshape_transpose we can make this case more generic.
For example if in place of mul there's transpose followed by add. In this case we can flip the transpose and the add for the slice section 1. We can horizontally fuse the add for the slices and then transpose everything.

@15 = gpu::code_object[code_object=7632,symbol_name=mlir_dot_add,global=133632,local=256,](@13,@12,@5,@14) -> half_type, {24, 77, 2304}, {177408, 2304, 1}: 0.0934304ms, 2%
@16 = hip::hip_copy_literal[id=main:@literal:78] -> half_type, {768}, {1}: 0.00109522ms, 1%
@17 = hip::hip_copy_literal[id=main:@literal:59] -> half_type, {768}, {1}: 0.00108192ms, 1%
@18 = slice[axes={2},starts={768},ends={1536}](@15) -> half_type, {24, 77, 768}, {177408, 2304, 1}: 0.00165542ms, 1%
@19 = multibroadcast[out_lens={24, 77, 768},out_dyn_dims={}](@17) -> half_type, {24, 77, 768}, {0, 0, 1}: 0.00094074ms, 1%
@20 = load[offset=18184320,end=21022848](@1) -> half_type, {24, 77, 768}, {59136, 768, 1}: 0.00076536ms, 1%
**@21 = gpu::code_object[code_object=5128,symbol_name=add_kernel,global=354816,local=1024,](@19,@18,@20) -> half_type, {24, 77, 768}, {59136, 768, 1}: 0.0211362ms, 1%**
@22 = load[offset=11354112,end=14192640](@1) -> half_type, {24, 77, 768}, {59136, 768, 1}: 0.00099472ms, 1%
@23 = multibroadcast[out_lens={24, 77, 768},out_dyn_dims={}](@16) -> half_type, {24, 77, 768}, {0, 0, 1}: 0.00182424ms, 1%
@24 = slice[axes={2},starts={0},ends={768}](@15) -> half_type, {24, 77, 768}, {177408, 2304, 1}: 0.00103286ms, 1%
**@25 = gpu::code_object[code_object=5136,symbol_name=mul_add_kernel,global=354816,local=1024,](@24,@23,@22) -> half_type, {24, 77, 768}, {59136, 768, 1}: 0.0413997ms, 1%**
@26 = load[offset=14769216,end=18184320](@1) -> half_type, {24, 12, 77, 77}, {71148, 5929, 77, 1}: 0.00105ms, 1%
@27 = gpu::code_object[code_object=6736,symbol_name=mlir_reshape_transpose_reshape_transpose_dot,global=73728,local=256,](@25,@21,@26) -> half_type, {24, 12, 77, 77}, {71148, 5929, 77, 1}: 0.0248955ms, 1%
...
@32 = load[offset=14769216,end=17607744](@1) -> half_type, {24, 77, 768}, {59136, 768, 1}
@33 = multibroadcast[out_lens={24, 77, 768},out_dyn_dims={}](@31) -> half_type, {24, 77, 768}, {0, 0, 1}
@34 = slice[axes={2},starts={1536},ends={2304}](@14) -> half_type, {24, 77, 768}, {177408, 2304, 1}
**@35 = gpu::code_object[code_object=5128,symbol_name=add_kernel,global=354816,local=1024,](@33,@34,@32) -> half_type, {24, 77, 768}, {59136, 768, 1}**

Start at the simplify_algebra entry point and trace how mul_add_kernel, add_kernel, and mlir_dot_add are currently horizontally fused. Use the provided graph as the behavioral example; done means the fusion is generalized to the described transpose-followed-by-add case and produces the intended combined operations.