Exclude const eval layer from flattening logic

[jazpurTT]

On TTIR + TTNN graphs we separate all the const eval graphs into their own layer called "const_eval" for better visualization. This request comes from #5822, but when we flatten the graph we loose this layer grouping and the rendering of the graph get's complex. We want to be able to exclude the "const_eval" layer from the flattening logic in model-explorer.

[jazpurTT]

Adding more context, the TT-NN MLIR modules we generate have a number of sub-graphs/functions in them that are used evaluate constant values and then get loaded into the main model.

ttcore.device_module {
    builtin.module @SyncTensorsGraph.4326 attributes { ... } {
      func.func private @main_const_eval_0() -> tensor<8x1x1x384xf32, #ttnn_layout> attributes {const_eval} {
        %0 = "ttnn.get_device"() <{mesh_offset = #ttnn<mesh_offset 0x0>, mesh_shape = #ttnn<mesh_shape 1x1>}> : () -> !ttnn.device loc(#loc)
        %1 = "ttnn.full"(%0) <{dtype = #ttcore.supportedDataTypes<f32>, fill_value = -3.38953139E+38 : f32, layout = #ttnn.layout<tile>, shape = #ttnn.shape<8x1x1x384>}> : (!ttnn.device) -> tensor<8x1x1x384xf32, #ttnn_layout> loc(#loc)
        return %1 : tensor<8x1x1x384xf32, #ttnn_layout> loc(#loc)
      } loc(#loc)
      func.func @main( ... ) {
        %0 = ttcore.load_cached(@main_const_eval_0, []) : () -> tensor<8x1x1x384xf32, #ttnn_layout> loc(#loc)
}

In medium to large models we have seen 100+ const-eval graphs and they are not overly informative when it comes to the main model so we were asked to render "out" the const eval models. The solution we came up with was to generate a separate layer containing all the cons-eval layer in their own sub-layer:

issue-const-eval.mov

But, when we flatten the graph we loose this layer grouping and the rendering of the graph get's complex and messy again:

issue-flatten.mov

We want the UI/front-end to to exclude the "const_eval" layer from the flattening logic in model-explorer. so the @main model get's flattened but all the const eval functions stays compressed in it's own const eval layer and sub-layers.

MLIR file from the above examples:

ttnn_1768020749900.txt