Fix: inject LoRA adapters before loading LoRA weights in from_pretrained

[MarkovChain-why]

Summary

• Bug: When defer_lora_injection=True, VLA.from_pretrained() (inference path) calls load_lora_weight() without first injecting LoRA adapters. The LoRA weight keys (lora_A, lora_B) have no matching parameters, so loaded weights are silently dropped.
• Root cause: The else branch correctly calls inject_lora_after_loading(), but the if lora_weights_path is not None branch does not.
• Fix: Call inject_lora_after_loading() before load_lora_weight() when a LoRA weights path is provided.

Note: The training path in base.py (create_model) is not affected — it already calls inject_lora_after_loading() unconditionally after loading pretrained weights.

Affected code

groot/vla/model/dreamzero/base_vla.py, VLA.from_pretrained() method.

Before (buggy):
```python
if lora_weights_path is not None:
model.load_lora_weight(lora_weights_path) # LoRA adapters don't exist yet!
else:
if ... defer_lora_injection:
model.action_head.inject_lora_after_loading()
```

After (fixed):
```python
if lora_weights_path is not None:
if hasattr(model.action_head, 'inject_lora_after_loading'):
model.action_head.inject_lora_after_loading() # Create adapters first
model.load_lora_weight(lora_weights_path) # Then load weights
else:
if ... defer_lora_injection:
model.action_head.inject_lora_after_loading()
```

Test plan

• Load a LoRA-finetuned checkpoint with defer_lora_injection=True via VLA.from_pretrained() and verify LoRA weights are correctly applied
• Verify inference outputs differ from base model (confirming LoRA weights took effect)