Add DecomposeGruPass for ARM backend (#17137)

Summary:

Adds a decomposition pass that transforms aten.gru.input into elementary
ops supported by TOSA (matmul, sigmoid, tanh, mul, add, slice, cat).

GRU cell equations per timestep:
    r_t = sigmoid(x_t @ W_ir.T + b_ir + h_{t-1} @ W_hr.T + b_hr)
    z_t = sigmoid(x_t @ W_iz.T + b_iz + h_{t-1} @ W_hz.T + b_hz)
    n_t = tanh(x_t @ W_in.T + b_in + r_t * (h_{t-1} @ W_hn.T + b_hn))
    h_t = n_t + z_t * (h_{t-1} - n_t)

Features:
- Multi-layer GRU support
- Bidirectional GRU support
- With/without bias
- batch_first support
- Batched gate computation (2 mm ops per timestep instead of 6)

Differential Revision: D92058313

Author: Andrew Pullin

backends/arm/_passes/decompose_gru_pass.py

@@ -66,6 +66,9 @@ def _build_direction(
     ) -> Tuple[List[torch.fx.Node], torch.fx.Node]:
         """Build GRU cell computation for one direction.
 
+        Uses aten.linear (matching PyTorch's standard decomposition) instead
+        of raw mm to avoid ConvertMmToBmmPass issues with quantized tensors.
+
         Returns (timestep_outputs, h_final) where timestep_outputs are
         unsqueezed hidden states in forward time order.