Add fallback mirrors to dataset API

[abidwael]

To mitigate the flakiness of external APIs and downtime from the main mirrors, this PR introduces fallback mirrors to the dataset API. These mirrors are defined in the dataset configs under ludwig/datasets/configs by adding

fallback_mirrors:
  - name: mirror_1
    download_paths:
      - path/to/dataset/file_1.json
      - path/to/dataset/file_2.json
  - name: mirror_2
    download_paths: 
     - ...

Only paths to filesystems are supported.

When downloading the dataset, the loader will try to retrieve the dataset from the main source (download_urls or kaggle_competition). If any exception is encountered, it will attempt to load the dataset from the mirror in the order specified.

[github-actions]

Unit Test Results

         6 files  ±0           6 suites  ±0   5h 37m 59s ⏱️ - 12m 22s
  3 983 tests +3    3 939 ✔️ +3    44 💤 ±0  0 ❌ ±0 
11 970 runs  +9  11 829 ✔️ +9  141 💤 ±0  0 ❌ ±0 

Results for commit 771d420. ± Comparison against base commit 04efab4.

♻️ This comment has been updated with latest results.