feat(optim): add algebraic graph rewrites

[xieofxie]

Summary

• add an opt-in algebraic rewrite pipe after ORT constant folding
• replace statically bounded Split nodes with Slice nodes
• fold safe channel-wise affine branches into producing Conv weights and bias
• fold inference Conv + static Add + BatchNormalization patterns
• guard graph outputs, nested subgraph captures, shared consumers, dynamic shapes, and repeated optimizer runs
• add generated ONNX numerical-equivalence and graph-safety coverage

QNN result

On the evaluated fixed-batch U-Net, the combined rewrites plus htp_graph_finalization_optimization_mode=3 reduced p50 latency from 31.125 ms to 22.729 ms and reduced QNN DDR traffic by roughly 30%.