[Feat] Add Qwen3-VL backbone support

- Add Qwen3-VL VLM backbone and register it in the VLM configs/factory
- Add GR00T + Qwen3-VL LIBERO finetune configs (0.6B spatial/object/goal/10, 2B all)
- Load the tokenizer from the checkpoint path during LIBERO eval
- Set trust_remote_code for pretrained tokenizer loading to silence custom-code warning
- Plumb backbone changes through condition_gemma / hf_vlm / paligemma / pi0_flowmatching
- Bump transformers to 5.3.0 in requirements

[Feat] Add Qwen3-VL backbone support

- Add Qwen3-VL VLM backbone and register it in the VLM configs/factory
- Add GR00T + Qwen3-VL LIBERO finetune configs (0.6B spatial/object/goal/10, 2B all)
- Load the tokenizer from the checkpoint path during LIBERO eval
- Set trust_remote_code for pretrained tokenizer loading to silence custom-code warning
- Plumb backbone changes through condition_gemma / hf_vlm / paligemma / pi0_flowmatching
- Bump transformers to 5.3.0 in requirements

[Feat] Support training-time RTC for PI0.5 and configurable delay tem…

…perature

    - Enable training-time RTC on PI05FlowMatching by removing the ValueError guards and making GemmaRMSNorm accept per-position adarms_cond of shape (B, L, cond_dim) in addition to (B, cond_dim)
    - Add temperature knob to sample_training_delay: scales the exponential delay weights (1.0 matches original behavior; higher values flatten toward uniform)
    - Update adarms_cond docstrings across GemmaForCausalLM / GemmaForSequenceClassification / GemmaForTokenClassification to reflect the new (B, L, cond_dim) shape
    - Update docs/rtc.md and scripts/test_rtc.py: document PI0.5 training-time RTC support, the temperature parameter, and drop the previous "PI0.5 unsupported" notes and skip-prefix branch