[Task] text-to-3d model support

Overview

Text-to-3D models generate 3D geometry directly from a natural language description. TRELLIS supports text conditioning via the same SLAT pipeline as its image variant; shap-e (OpenAI) generates implicit neural representations from text; BrickGPT specializes in generating LEGO-like brick structures from text prompts.

Agent Scenarios

• Creative content agent: generate 3D scene elements (characters, props, environments) from a designer's text description for prototyping
• Game / metaverse asset agent: instantly produce game-ready geometry from narrative prompts without a 3D artist in the loop
• Education / toy design agent: use BrickGPT to generate brick-by-brick building instructions from a text description of a structure
• Concept visualization agent: turn a product brief into a rough 3D mockup for early-stage design review

ModelKit Integration

Models must pass the full wmk pipeline on all EPs:

wmk config → wmk build (ONNX export) → wmk perf → wmk eval

Acceptance Criteria

• microsoft/TRELLIS-text-xlarge
• microsoft/TRELLIS-text-large
• openai/shap-e
• AvaLovelace/BrickGPT

Start with the ModelKit integration flow described in the issue: run wmk config, wmk build for ONNX export, wmk perf, and wmk eval. Work through the four listed models—TRELLIS-text-xlarge, TRELLIS-text-large, shap-e, and BrickGPT—and consider the task complete when each passes the full wmk pipeline on all EPs.