[QNN EP] Handle broadcast Gemm C with FC and Add

[qti-yuduo]

Summary

• Lower valid Gemm C broadcasts that cannot be represented as a QNN FullyConnected bias to FullyConnected + ElementWiseAdd.
• Share the FC-compatible bias-shape rule and keep the absorbed-Reshape selector aligned with the builder.
• Add direct, MatMulAddFusion, and selector regression coverage.

Root cause

ORT MatMulAddFusion can produce Gemm with C=[M,1] from MatMul followed by a column-broadcast Add. ONNX permits C to broadcast to [M,N], but QNN FullyConnected can consume only one bias value per output channel. Validation rejected C before the existing FC + Add lowering could run.

The fix keeps scalar C for N=1, [N], and [1,N] on the direct FC path. Other valid broadcasts use a separate Add; block-quantized Gemm remains strict.

[qti-ashwshan]

LGTM! we are correctly gating ExplictOpCheck to let decomposable bias shapes flow through to the existing FC+Add path instead of rejecting them outright.