Fix H0.5 benchmark inference compatibility

[MarkfuGod]

Summary

• Fix the get_modality_config inference endpoint so it is invoked without a request payload.
• Add a generic BaseDataConfig.modality_config() implementation for client-side modality discovery.
• Cast LIBERO and RoboCasa benchmark state observations to float32 before sending them to the H0.5 inference server.

Why

While running the released Being-H0.5 LIBERO inference server and benchmark client end-to-end, two server-side compatibility issues appeared:

• get_modality_config was registered as a payload-taking endpoint, causing an extra argument to be passed into BeingHPolicy.get_modality_config().
• BeingHPolicy.get_modality_config() expected data configs to expose modality_config(), but BaseDataConfig did not provide it.

The LIBERO benchmark then reached a real rollout but failed during state preprocessing because simulator observations were sent as float64, while the H0.5 transform pipeline expects consistent float32 state tensors. RoboCasa uses a similar observation path, so this patch applies the same dtype normalization there.

Test plan

• python -m py_compile Being-H05/BeingH/inference/beingh_service.py Being-H05/configs/data_config.py Being-H05/BeingH/benchmark/utils/policy.py Being-H05/BeingH/benchmark/robocasa/run_robocasa_eval_fast.py
• Loaded Being-H05-2B_libero with BeingHPolicy.
• Started BeingH.inference.run_server_vla locally.
• Verified client ping.
• Verified client get_modality_config.
• Verified remote dummy get_action.
• Ran minimal LIBERO rollout:
  • suite: libero_spatial
  • task id: 0
  • trials: 1
  • result: 1/1 success
• Ran full LIBERO evaluation with robosuite==1.4.0:
  • libero_spatial: 493/500, 98.6%
  • libero_object: 485/500, 97.0%
  • libero_goal: 490/500, 98.0%
  • libero_10: 475/500, 95.0%