Add If rank alignment and LSTM rank reduction to enable silero_vad conversion

[evkotov]

Details:

This PR adds rank alignment for ONNX If nodes and rank reduction for LSTM operations to enable conversion of models like silero_vad.onnx.

Problem 1: If Node Rank Mismatch

ONNX If nodes with branches producing outputs of different ranks were causing dynamic rank in the output, which prevented further graph optimizations and caused runtime errors in operations requiring static rank (Conv, ReduceMean).

Example:

If (condition) {
  then_branch: ReduceMean(x, axes=[1], keepdims=1) → output shape [2, 1]
  else_branch: Identity(x) → output shape [2, 4]
}
Result: Dynamic rank output [?, ?]

Before fix:

   If Node
      |
     ┌────┴────┐
     |         |
     v         v
  then: [2,1] else: [2,4]
     |         |
     └────┬────┘
          v
     Output: [?, ?]  (dynamic rank)
          v
     Conv/ReduceMean
          v
     ERROR: "rank must be static"

After fix:

   If Node
      |
     ┌────┴────┐
     |         |
     v         v
  then: [2,1] else: [2,4]
     |         |
     v         |
  Unsqueeze   |
  (axis=0)    |
     |         |
  [1,2,1]  [1,2,4]
     |         |
     └────┬────┘
          v
     Output: [1,2,?]  (static rank 3)
          v
     Conv/ReduceMean
          v
     SUCCESS

Problem 2: LSTM High-Rank Inputs

Models like silero_vad.onnx have LSTM layers where input tensors have rank > 3 (e.g., [1,1,?,?,?]), but OpenVINO LSTM expects exactly rank 3 [batch, seq, features].

Example from silero_vad:

X: [1, 1, ?, ?, ?]  (rank 5)
initial_h: [1, 1, 2, 1, 64]  (rank 5)
initial_c: [1, 1, 2, 1, 64]  (rank 5)

Before fix:

  X [1, 1, ?, ?, ?]
        |
        v
     ┌──┴──────────────┐
     | LSTM Operator   |
     | Expects rank=3  |
     | Got rank=5      |
     └─────────────────┘
  Error: "Shape rank of input at 0 is incompatible.
          Expected rank: 3, actual shape: [1,1,?,?,?]"

After fix:

  X [1, 1, ?, ?, ?]
        |
        v
  reduce_tensor_rank(target=3)
        |
        v
  Squeeze leading dims = 1
   or Reshape if needed
        |
        v
  X [batch, seq, features]  (rank 3)
        |
        v
     ┌──┴──────────────┐
     | LSTM Operator   |
     |   rank=3       |
     └─────────────────┘

Tickets:

• 162986

[mvafin]

You comminted .onnx models instead of .prototxt

[mvafin]

@evkotov Not very clear why we need this change for If, what is the reason for rank mismatch? Why we add Unsqueeze? Can this be done in a separate PR?

[evkotov]

You comminted .onnx models instead of .prototxt

I updated it

[evkotov]

moved LSTM fix part to #33023

[evkotov]

moved if fix to #33027