Add `ul cloud train|predict|export` workflows (#71) * Add ul cloud train, predict and export workflows Run YOLO training, inference and export on Ultralytics Platform with the same key=value arguments as the local commands. Local weights and datasets are uploaded, jobs are awaited, and outputs are saved locally with YOLO's own writers. * Handle optional predict metadata, archive-relative data.yaml, per-call Platform URL Address review: default class names when classNames is absent and an explicit error when task is absent, drop the local path: key from the archived dataset YAML, and read ULTRALYTICS_PLATFORM_URL per invocation. * Name the per-image prediction parameter and document name= as the output model name * Prefer local checkpoints over same-named official weights, resolve model stems, let Platform infer dataset tasks Address review: an existing local file is uploaded even when it shares a name with official weights, model stems resolve to .pt as in YOLO, and datasets are created without a guessed task because ingest infers it from the labels. * Restrict cloud export downloads to the output directory * Handle missing cloud export download metadata * Validate dataset split paths before cloud packaging * Allow local inference settings in cloud prediction * Warn and ignore local-only cloud export options * Package validation-alias, relative-path, and images-rooted datasets; tolerate empty masks; validate local args before training * Report an outdated ultralytics install as a controlled cloud workflow error * Map official weights to public Platform models only when they are hosted * Keep ul:// model URIs intact when normalizing model stems * Treat null cfg values as unset so copy-cfg YAMLs keep the workflow defaults * Archive split paths relative to the dataset root, tolerate a missing ingest status, and report upload failures * Resolve '../' split values under the dataset root as check_det_dataset does --------- Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>