feat: Add ESMFold protein structure prediction job bundle

[folouiseAWS]

Adds a Deadline Cloud sample for protein folding via Meta's ESMFold v1 (facebook/esmfold_v1, MIT). Demonstrates a 4-step splitter-then-fanout pipeline pattern (SplitFasta -> Fold -> Render -> Validate) for variable- count GPU ML inference workloads on a service-managed fleet.

The bundle:

• Parses a user FASTA, validates sequences, round-robins them across N GPU tasks longest-first for balanced runtimes
• Runs ESMFold inference with three layers of validation (structural sanity, mean pLDDT self-consistency, and optional TM-score / RMSD / pLDDT calibration vs experimental references)
• Renders matplotlib CA-trace structures colored by per-residue pLDDT
• Produces a per-sequence calibration plot showing model confidence vs actual error against ground-truth structures (when refs are provided)

Reuses the splitter manifest pattern from copy_s3_prefix_to_job_attachments and the conda queue-environment pattern shared across ML bundles.

Fixes: N/A — new sample, not tied to a tracked issue

What was the problem/requirement? (What/Why)

The samples repo had ML/GPU bundles for image generation (flux2_klein_lora) and 3D reconstruction (gsplat_pipeline) but no example demonstrating the splitter-then-fanout
pattern for variable-count GPU inference workloads. Protein structure prediction is a clean fit: the input is a FASTA file with N records (where N is unknown until the

1. SplitFasta (1 task, CPU): parses the input FASTA, validates each sequence (length and amino-acid alphabet), and round-robins the records across Parallelism per-task
   manifest files in a shared workspace directory. Sorts longest-first so per-task runtimes balance.
2. Fold (Parallelism tasks, GPU): loads facebook/esmfold_v1 from HuggingFace, runs inference on each sequence in the assigned manifest, validates the output PDB (atom
   count, NaN check, pLDDT range), and emits one .pdb plus a summary.json per sequence. Uses the same HF cache pattern (HF_HOME pointed at OutputDir/.hf_cache) as other ML
   bundles.
3. Render (1 task, CPU): produces a matplotlib backbone trace per predicted structure, colored by per-residue pLDDT confidence.
4. Validate (1 task, CPU, optional): when a ReferencePdbDir is provided, computes TM-score (via biotite's native tm_score), RMSD, and a per-residue pLDDT/error Pearson
   correlation against each experimental reference. Writes a validation.csv and a per-sequence calibration.png.

The splitter + manifest pass-through pattern is borrowed from copy_s3_prefix_to_job_attachments. The conda-queue-environment pattern is shared with the other GPU/ML
bundles.

What is the impact of this change?

• Adds a new Deadline Cloud sample for scientific GPU compute with no impact on existing bundles.
• All inputs/outputs go through standard job attachments; no new IAM permissions required.
• The bundle's CondaPackages resolves cleanly against the same cuda_farm CloudFormation farm template the other GPU samples target.

How was this change tested?

Submitted to a cuda_farm queue (g5 fleet, A10G workers) and iterated through real failures and fixes:

• Verified end-to-end pipeline. Folded the demo set (1L2Y, 2JOF, 1VII) on the farm. All four steps green:
  SplitFasta → SUCCEEDED
  Fold → SUCCEEDED (5.12 s for 1VII, ~1 s for trp-cages)
  Render → SUCCEEDED (3 PNGs produced)
  Validate → SUCCEEDED (validation.csv with TM-score + Pearson r)
• Verified validation against ground truth. Fetched experimental PDBs from RCSB for the three demo sequences, ran with ReferencePdbDir set, confirmed:
  seq_id,tm_score,rmsd,aligned_residues,mean_plddt,plddt_error_pearson
  1l2y,0.5589,0.549,20,85.54,-0.2767
  1vii,0.6555,2.170,36,86.90,-0.5888
  2jof,0.4530,0.967,20,87.25,-0.2876
• RMSD < 1 Å for both trp-cages indicates correct prediction; Pearson r negative for all three confirms model confidence is calibrated.
• Visually verified the rendered output by downloading the predicted structures and comparing CA-traces against experimental references. The two images embedded in the
  README (1vii_structure.png, 1vii_calibration.png) are real outputs from a successful farm run, copied into .readme_images/ for documentation.
• Iterated through real bugs during testing and fixed each: a WorkspacePath: dataFlow issue blocking manifest pass-through, a pLDDT scale conversion ([0,1] → [0,100])
  needed for AlphaFold-convention rendering, and a switch from a custom TM-score implementation to biotite's native one for correctness on alignment-aware residue matching.

Was this change documented?

Yes — README.md covers prerequisites (farm setup, VRAM table, service quotas), submission examples (smoke test, with-validation, custom inputs), parameter reference,
output layout, validation strategy (three layers including the calibration plot), and notes on weight caching and monomer-only scope. sample_inputs/README.md documents
the demo FASTA contents and how to fetch reference PDBs from RCSB.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[leon-li-inspire]

This import has not been used.

[crowecawcaw]

Nit: better to put at top of file if possible

[crowecawcaw]

Nit: put this version requirement in the conda packages string

[crowecawcaw]

Nit: put these imports at top of file if possible, same below