[Enterprise] Add end-to-end Boltz-2 support

[yifeif-nv]

Parent epic: #998

Summary

Add end-to-end TensorRT Model Connect support for a pinned Boltz-2 structure-prediction workflow.

The implementation must cover the complete model path and the data transformations required to turn documented biomolecular inputs into validated structure and confidence outputs.

Tasks

• Pin the exact Boltz-2 source repository, source revision, checkpoint revision, configuration, and applicable licenses.
• Define supported sequence, MSA, template, ligand, constraint, and feature inputs.
• Define the output structure, confidence, ranking, and metadata contracts.
• Implement the Boltz-2 model graph and model-owned input/output transformations.
• Support the required Pairformer or trunk, recycling, diffusion, masking, and variable-length behavior.
• Package all required inference components and metadata into a reproducible Model Connect bundle.
• Add a native runtime example that does not require Python or PyTorch after bundle creation.
• Validate structure and confidence outputs against the pinned reference implementation using documented quality metrics and tolerances.
• Establish an initial BF16 qualification profile on at least one documented supported NVIDIA GPU.
• Benchmark warmed steady-state latency, throughput where applicable, and peak memory against aligned eager PyTorch and torch.compile baselines, excluding engine-build time.
• Record valid-output rate, unsupported inputs, maximum tested sizes, and known limitations.
• Add model-owned unit, integration, qualification, and end-to-end tests.
• Document clean-environment setup, bundle generation, native execution, and result interpretation.

Acceptance criteria

A pinned Boltz-2 configuration processes a documented example end to end through a native TensorRT Model Connect bundle and produces structurally valid outputs within the agreed quality tolerances.

The qualified BF16 profile records the exact GPU and software stack and includes reproducible quality, validity, latency, and memory evidence plus documented supported inputs and limitations.