RibonanzaNet-SS, folding util extraction

[luxaritas]

Summary

• Add support for the RibonanzaNet-SS folding engine
• Extract some generic folding utilities (pUnpaired, threshknot) into FoldUtil for easier reuse (and readability)

Implementation Notes

• RibonanzaNet-SS inference setup was based on https://www.kaggle.com/code/shujun717/ribonanzanet-2d-structure-inference/
• We've introduced a new asset type, onnx (a standard format for ML models). Particularly notable is that our jest setup has to actually have imports of this asset provide its path so that it can be loaded to run in tests. I've changed all generic assets to do this (instead of just being an empty module) in case something else in the future requires the assets to be actually loaded.
• onnxruntime (responsible for doing the actual model inference) can only be run async, so we've had to ensure all codepaths which trigger folding operations can operate asynchronously (or, in the case of scripts, error when attempting to run synchronously). We use the existing PoseOp mechanism (originally built for allowing multistrand folding to yield to the main thread intermittently when not run from scripts) to do this, with some slight modifications.
• We have introduced new sometimes-sync-sometimes-async codepaths with function overloads, where a boolean determines whether it is marked async and returns a promise or is not marked async and returns the value directly. The actual implementation is always an async function, however the JS spec specifies that if no await is ever encountered, the function will run synchronously.
• When dot plot calculations are needed for constraints, instead of triggering the calculation in the constraint evaluate function (which has to run synchronously), we set a property on the constraint so that we can queue the calculation during the folding process (so that the data is already available when evaluate is run). We've also introduced the skipMelt parameter to updateMeltingPointAndDotPlot - this way if we're only running this update for the constraints, we don't have to run the extra fold/dot plot operations needed for the melt plot (which no constraints use)
• Unfortunately we can't make folding multithreaded, even though onnxruntime supports it (and we'd potentially gain something like a 2x speedup), as that requires a cross origin isolated browsing context, which we can't set up because of YouTube and Vimeo embeds :(

Testing

• New unit tests
• Verified a couple samples were consistent with the kaggle notebook
• Misc interactive testing, including user feedback via the forums

[luxaritas]

TODOs

RNNet-SS Behavior

• Updated mask diagonal handling (for 2nt hairpins)
• Revisit higher-order pseudoknot limit (add letter-based pairs, error if cardinality is too high)
• Investigate quantization (Deferred - planning on investigating this + distillation with Shujun later)

Metrics

• Fix energies still showing in puzzlemaker (Consistently disable energies when unavailable #757)
• eF1 (+ eF1,crossed-pair)
• eF1 (+ eF1,crossed-pair) constraints (Deferred - don't need this right now, need to think through how this behaves with other engines)
• Investigate dot plot being largely gray instead of white
• Fix dot plots missing target? (Deferred - need to figure out how to handle the fact this has caused an inconsistency in the data exposed to EternaScript)

Engine implementation and access

• Change path to rnnet onnx file to engines-bin
• Lazy-load/initialize rnnet on first usage
• Async APIs for Eternascript
• Add tests (fold, dot plot, eF1)

Asset handling

• Look into enabling secure context to enable threading (EDIT: Looks like we're out of luck)
• Ensure wasm/onnx/ort files are cached
• Investigate loading bar (Deferred - lazy-loading is sufficient for now)
• Download confirmation? (Deferred - lazy-loading is sufficient for now)
• Investigate splitting out sourcemaps (Improve sourcemap behavior #760)
• Determine whether we need a "minimal" onnxruntime build (or maybe import from /wasm?) onnxruntime appears to be a relatively limited portion of our bundle size. I tried importing from /wasm and it saved like 300k, but that's not really that meaningful in context, so I'm leaving it as-is so that we won't have to change it back later if we can eventually enable GPU acceleration.

[luxaritas]

Make sure to note in PR description:

• Async functions without await run synchronously
• Explain why we introduced skipMelt

[tkaragianes]

Reviewed in sync conversation with Jonathan, notes are in his review.

[tkaragianes]

Closes #788.