fix: add guidance_scale, distilled config and latents tuple compat

[liz-badada]

Summary

• Add guidance_scale parameter to GenerateRequest and propagate it through request building and cache metadata.
• Set a sensible DEFAULT_NEGATIVE_PROMPT to improve generation quality out of the box.
• Force Distilled-mode pipeline config (is_distilled, is_enable_stage2, is_amplify_first_chunk, pyramid_num_inference_steps_list) during worker init, since sglang may not auto-detect these when HeliosDMDScheduler is absent from its registry.
• Fix compatibility issue where some sglang versions return (latents, ...) tuple instead of a plain tensor in the patched denoising forward.

Files Changed

• examples/helios_continuation/helios_continuation_dynamo.py
• examples/helios_continuation/helios_latent_cache.py

Test

# Server Launch
./examples/helios_continuation/run_helios_continuation.sh


# Step1
curl -X POST http://localhost:8090/v1/videos/generate \
  -H 'Content-Type: application/json' \
  -d '{"prompt": "A cat walking on grass", "num_frames": 33, "height": 384, "width": 640}'

# Step2
curl -X POST http://localhost:8090/v1/videos/continue \
  -H 'Content-Type: application/json' \
  -d '{"cache_id": "<cache_id returned by step1>", "num_frames": 33}'