Clarify 32D base action space padding

[taivu1998]

Summary

• Document dimensions dim_16:dim_31 as reserved padding in the provided base action-space convention.
• Replace the partial code-block action layout with complete tables covering the 32D interface and common robot adapter slices.
• Clarify that the 32D width is a model interface detail: adapters pad shorter vectors before the model and slice undefined dimensions away before robot execution.
• Note that some pi05_base fine-tuning configs load compatible normalization stats from the pi0.5 base assets, without overclaiming every listed asset is available for every checkpoint.

Root Cause

docs/norm_stats.md described the leading robot action dimensions through dim_15, but the released base model interface can expose a 32D action width. Without an explicit explanation of dim_16:dim_31, users could reasonably infer those dimensions encode hidden robot controls.

User Impact

This makes the action-space documentation self-contained for users interpreting model outputs or wiring custom robot adapters. It makes clear that undefined tail dimensions should not be sent to robot controllers unless a custom adapter, dataset, and matching normalization stats define them.

Validation

• Reviewed source behavior for PadStatesAndActions, normalization/unnormalization, and ALOHA/DROID/LIBERO output slicing.
• Ran git diff --check.
• Ran targeted rg checks for the new dim_16:dim_31 docs and source references.
• Attempted focused pytest for transform helpers, but local macOS arm64 setup cannot install jax-cuda12-plugin==0.5.3 because it only ships Linux wheels.

Closes #748