Add build-custom-learning-blocks experimental skill

[yennster]

Summary

Adds skills/.experimental/build-custom-learning-blocks, an authoring guide for Edge Impulse custom learning blocks (custom ML blocks), plus its row in the README experimental catalog table.

Source

Distilled from the official documentation and example repositories:

• https://docs.edgeimpulse.com/studio/organizations/custom-blocks/custom-learning-blocks (plus the custom-blocks overview and the parameters-json / train-input-json / sample-id-details-json specs and the blocks CLI reference)
• https://github.com/edgeimpulse/example-custom-ml-block-keras
• https://github.com/edgeimpulse/example-custom-ml-block-pytorch
• https://github.com/edgeimpulse/ml-block-yolov5

All linked repositories confirmed public and not archived; all linked docs URLs confirmed resolving.

Intended use

Agents asked to scaffold, modify, test, or push a custom learning block: the training-container CLI contract, data directory formats (classification, object detection, anomaly), output artifact requirements, parameters.json for machine-learning blocks, Dockerfile rules, Keras/PyTorch train-script skeletons, local testing with edge-impulse-blocks runner, and the init/push workflow.

Overlap analysis

Complements the stable build-custom-deployment-blocks skill (deployment blocks: deployment-metadata.json in, deploy.zip out) — no overlap in contract; the two block types share only the edge-impulse-blocks CLI. No other skill covers training containers. The firmware-* skills consume exported models and sit downstream of this one.

Validation performed

• skills-ref validate passes at the same pinned commit CI uses (38a2ff8); git diff --check clean.
• Claims were cross-checked against the three example repos (exact CLI flags, filenames, code) and the live docs pages; genuine upstream inconsistencies are flagged in-place in the skill (--info-file vs --input-file, one-hot vs structured Y arrays, commit-vs-gitignore guidance for .ei-block-config).