[MAX] Implement Z-Image ModuleV2 pipeline by byungchul-sqzb · Pull Request #6330 · modular/modular

@byungchul-sqzb

Port Z-Image to the Graph API / ModuleV2 runtime using V2 text encoder,
transformer, and VAE components.

Restore the current ModuleV3 feature set and behavior in the V2 path,
including:
- Z-Image transformer/model/config/weight adapter wiring
- diffusion pipeline, arch registration, and ModuleV2/ModuleV3 selection via
  --prefer-module-v3
- batched CFG, CFG renormalization, and image parity with ModuleV3
- transformer-side RoPE micro-optimizations:
  - single unified RoPE embedder call
  - interleaved [cos, sin] frequency generation
  - rope_ragged_with_position_ids hot path
  - preamble dtype cast and direct modulation slicing