LayerNormFusion should reject shape-expanding scale and bias patterns

LayerNormFusion::ApplyImpl selects scale and bias primarily by rank and only compares their concrete dim_value() fields. It does not verify that the trailing Mul/Add preserve the normalized input shape.

As noted during review of #32058, an expanding pattern such as Mul([1, 1], [2]) + [2] can therefore be replaced with LayerNormalization, which preserves the input shape and changes semantics or fails at runtime. Different symbolic dimensions can also compare equal because dim_value() is zero when only dim_param is present.

Follow-up work:

• Select scale and bias by graph connectivity rather than rank alone.
• Reject fusion when the trailing broadcast is known to expand the normalized tensor.
• Compare symbolic dimensions correctly.
• Add positive operand-order and negative expanding-broadcast regression tests.

Related: #25855 and the sibling SimplifiedLayerNormFusion fix in #32058.

Start at LayerNormFusion::ApplyImpl and compare it with the sibling SimplifiedLayerNormFusion fix in #32058. Trace scale and bias through the graph, then add positive operand-order and negative expanding-broadcast regression tests; done means symbolic dimensions are distinguished and invalid expansions no longer fuse.