fold-cp documentation and examples

[zyndagj]

Adds a BioNeMo Contessa cookbook for running ESMFold2 with single-GPU accelerated backends and context parallelism (Fold-CP).

This includes:

• New single-GPU examples for:
  • PyTorch reference backend (esmfold2-none.py)
  • cuEquivariance backend (esmfold2-cueq.py)
  • Triton fused backend (esmfold2-fused.py)
• New Fold-CP examples for:
  • CP setup on the same small 7ysz input (esmfold2-cp.py)
  • Larger 5xgo input that fits with CP (cuequivariance_cp.py)
  • Single-GPU OOM demonstration for 5xgo (will_fail.py)
• Documentation for:
  • Single-GPU backend tradeoffs in single-gpu.md
  • wrap_model_with_cp usage, requirements, and behavior in fold-cp.md
  • Cookbook overview in README.md
• Optional dependency extras in pyproject.toml:
  • esm[fused]
  • esm[cueq12]
  • esm[cueq13]
  • esm[fold-cp]