Vision dataset downloading

[rdondera-microsoft]

Support downloading vision + multimodal datasets in the dataset downloader component.

Implementation
It is not possible to directly save vision HF datasets to JSONL (as is done for NLP) because binary image data is not compatible with this setup. The solution was to keep the JSONL format but use a URL field that refers to image data saved to the default datastore. For each HF dataset, a specific adapter class extracts a label and a PIL image for a dataset instance and common code uploads the image, passes through the fields etc.

Extensions
New vision/multimodal datasets from HF can be added by adding new dataset adapters.

Tests
Ran the benchmarking pipeline e2e in the cloud: https://ml.azure.com/experiments/id/9f7237b4-3232-49d3-ae02-e07326242835?wsid=/subscriptions/dbd697c3-ef40-488f-83e6-5ad4dfb78f9b/resourcegroups/rdondera/providers/Microsoft.MachineLearningServices/workspaces/benchmarking&tid=72f988bf-86f1-41af-91ab-2d7cd011db47 runs "resisc45", "food101", "patch_camelyon", "gtsrb".

[github-actions]

Test Results for assets-test

250 tests   249 ✅  4h 37m 2s ⏱️
 11 suites    1 💤
 11 files      0 ❌

Results for commit b788660.

♻️ This comment has been updated with latest results.