Fix load_lora to rebuild post-trained LoRA checkpoints on their recorded pretrained base

[dineshmane]

Fixes #82

Problem

VLA.load_lora composes LoRA checkpoints onto the DiT referenced by
config.json's diffusion_model_pretrained_path. For the post-training
workflow (yam_training.sh / agibot_training.sh), that path is the raw
Wan2.1 dir (from the dit_version flag) — but training loaded the full
DreamZero-AgiBot checkpoint (pretrained_model_path) before injecting the
adapters, so the saved deltas are relative to the AgiBot base. The shipped
load path therefore silently serves raw Wan + deltas: the model loads and
runs, with the robot pretraining discarded.

Fix

load_lora now reads pretrained_model_path from the checkpoint's
experiment_cfg/conf.yaml (already saved by every training run):

• Present → replicate training's composition: instantiate with
  defer_lora_injection=True, skip_component_loading=True, load the base
  checkpoint shards (same loop as create_model in
  groot/vla/experiment/base.py), inject_lora_after_loading(), then overlay
  the checkpoint's deltas. If the base dir is missing, raise a clear
  FileNotFoundError instead of silently substituting a different base.
• Absent (original from-scratch/DROID workflow) → unchanged behavior.

Validation

1000-step yam LoRA fine-tune of DreamZero-AgiBot on
lerobot/aloha_static_screw_driver (rev v2.1, adapted to yam embodiment).
Open-loop replay of 150 frames (scripts/open_loop_yam.py, slices adapted to
the 14-dim ALOHA layout); first predicted action vs ground truth, MSE in rad²:

Load path	overall	left_gripper
load_lora before this PR (raw-Wan base)	0.00988	0.07602
offline merge onto AgiBot base (ground truth)	0.00100	0.00180
load_lora with this PR	0.00100	0.00180
AgiBot base, no fine-tune (reference)	0.00736	0.00170

The pre-PR path is worse than not fine-tuning at all; the patched path
matches the offline-merged ground truth exactly (bit-identical MSE to 16 significant digits) — the patched loader reconstructs the same weights as the training-time composition.

Hardware: 4× RTX PRO 6000 Blackwell, torch 2.8.0+cu129, flash-attn 2.8.3.post1.

Notes

• No change for DROID-workflow checkpoints (no pretrained_model_path in conf.yaml).
• An alternative/complementary fix — rewriting diffusion_model_pretrained_path
  in the saved config at train time — was left out to keep existing checkpoints
  fixable without retraining; happy to add it if maintainers prefer.