Add reproducible legacy container environment

[Jake Smith (jsmith13)]

Summary

• add a pinned, multi-stage Python 3.9 / PyTorch 1.9 / Fairseq container
• compile Graphormer's Cython preprocessing extension and verify the installed runtime
• add PyG 2.x dataset protocol adapters and document CPU, GPU, and single-host distributed use
• exclude checkpoints, datasets, and experiment artifacts from the build context

The runtime retains g++ because PyTorch 1.9 JIT-compiles Graphormer's scripted preprocessing kernels during training. The image is intended as a reproducibility environment for the existing Fairseq-based code rather than a modernization of the dependency stack.

Validation

• docker build --check .
• clean image build from the final source tree
• CPU import and molecular preprocessing verification
• CUDA verification across four RTX A6000 GPUs
• pretrained PCQM4Mv2 checkpoint inference
• single-GPU training, validation, and checkpoint smoke test
• two-GPU NCCL/DDP training and checkpoint smoke test