PR for MoveLinearPastEltwiseMul transformation

[shashwat1198]

PR Description:

This PR introduces a new transformation, MoveLinearPastEltwiseMul, which optimises the computation graph by moving linear multiplication operations past elementwise multiplication operations when possible.

• Pattern before transformation:
  [(x × A) × (y × B)]

• Pattern after transformation:
  [(x × y) × (A × B)]

Where (x) and (y) are dynamic inputs, and (A) and (B) are constant tensors.

Key Changes:

• The transformation identifies mul operations that can be merged and moves linear mul nodes past elementwise mul nodes in the graph.
• The transformation checks for mul operations with shared initializers and merges them, removing redundant nodes.
• The model is updated with new initializers after the transformation to ensure correctness.
• A post-transformation pass is performed to infer shapes for any updated nodes in the graph.

[fpjentzsch]

We already have this in FINN+ as the "MoveConstMulPastJoinMul" transformation.

Christoph (@iksnagreb) added it as part of the StreamlinePlus PR (eki-project#39), which doesn't yet have an equivalent in this repository I believe.