feat: support Cosmos3-Edge-Policy-DROID in robolab policy server

[ychao-nvidia]

Summary

Enables serving the new Cosmos3-Edge-Policy-DROID checkpoint through the robolab action policy server. The Edge policy was post-trained from Cosmos3-Edge (Nemotron-3-Dense-VL backbone) and differs from the existing Nano policy in its processor, model config, tokenizer export, and prompt formatting — this MR wires up each of those differences so the checkpoint runs end-to-end.

Changes

• data/generator/processors/__init__.py — Build the Edge processor from nvidia/Cosmos3-Edge-Reasoner. The policy checkpoint bundles a custom Cosmos3EdgeProcessor with no loadable processor code (no auto_map / remote
  code), so it can't be built via AutoProcessor. Sourcing it from the reasoner repo it was post-trained from mirrors how internal training sources the Nemotron-3-Dense-VL processor.
• inference/common/public_model_config.py — Register the Nemotron3DenseVLMoTConfig.from_json_file and Nemotron3DenseVLTextForCausalLM target aliases used by the Edge policy.
• scripts/action_policy_server_robolab.py — Add the Edge policy HF repo mapping, and a new format_prompt_as_json server arg that serves prompts as structured JSON to match training. When the transform emits ai_caption as a dict, it's JSON-encoded before serving.
• utils/checkpoint_db.py — Re-form the bare relative Wan2.2 VAE path exported by the Edge checkpoint into an s3://bucket/... registry key. The Edge conversion exports the VAE tokenizer with an empty bucket_name (Nano exports bucket_name="bucket"), so the raw relative vae_path reaches the downloader instead of a registry key; re-forming it lets the VAE resolve through the registry and auto-download from HF, matching Nano's path. Local files still take precedence.

[rickzw]

LGTM