feat(mixer): use timm mixer instead of custom mixer

[yhisaki]

Summary

This PR replaces the custom MLP-Mixer implementation with the standard MixerBlock provided by timm.

The scene encoders and turn-indicator trajectory encoder now use the same upstream implementation, reducing duplicated code and making the Mixer architecture easier to understand and maintain.

Changes

• Replace the custom diffusion_planner.model.module.mixer.MixerBlock with timm.models.mlp_mixer.MixerBlock.

• Remove the custom mixer.py module.

• Update the following encoders to use the timm Mixer interface:

  • EgoEncoder
  • NeighborEncoder
  • LaneEncoder
  • LineEncoder
  • TrajectoryEncoder

• Simplify encoder constructors by removing the custom tokens_mlp_dim and channels_mlp_dim arguments.

• Use each encoder's embedding dimension as the Mixer channel dimension and the corresponding sequence length as the token dimension.

• Use timm's Mlp implementation for single-token and auxiliary MLP encoders where applicable.

• Apply drop_path_rate through the residual DropPath mechanism provided by the timm Mixer block.

Motivation

The previous Mixer implementation duplicated functionality already available in timm and required project-specific constructor arguments.

Using the upstream implementation:

• reduces custom model code,
• standardizes the Mixer architecture,
• makes the expected token and channel dimensions clearer,
• provides conventional MLP-Mixer expansion ratios and stochastic-depth behavior, and
• makes future maintenance and architecture changes easier.

Behavioral changes

This is not a parameter-for-parameter replacement of the previous Mixer block.

The timm implementation uses its standard token- and channel-MLP expansion ratios and applies drop_path_rate as stochastic depth on the residual branches. Therefore, the internal parameter shapes and regularization behavior differ from those of the previous custom implementation.

[SakodaShintaro]

Since retraining is required, it is likely to take some time.