fix: preserve timestep rank in sinusoidal embedding

[KakaruHayate]

Summary

• preserve the input rank in SinusoidalPosEmb
• restore the singleton frame axis after timestep MLPs when tracing rank-1 timesteps
• keep rank-2 dual-timestep training and interpolation behavior unchanged

Motivation

The ONNX exporters trace diffusion backbones with a rank-1 timestep tensor. Forcing that tensor to rank 2 inside SinusoidalPosEmb makes the timestep MLP operate on rank-3 inputs, which exports its linear layers as MatMul + Add. ONNX Runtime DirectML graph fusion rejects the resulting diffusion loop graph.

Keeping rank-1 timesteps rank 1 lets the linear layers export as 2-D Gemm, then the backbone restores the existing [B, 1, C] contract. Rank-2 dual timesteps already produce [B, 1, C] and are unchanged.

Validation

• rank-1 and rank-2 timestep paths produce identical [B, 1, C] output (max_abs_error = 0)
• legacy ONNX export emits two Gemm nodes and no MatMul for the timestep MLP
• ONNX checker passes