Set saved_weights_in_checkpoint immediately after creating model. Also adds test.

[dantreiman]

This is a follow-up to "Prevent loading pre-trained encoder weights when loading a trained Ludwig model".

The original issue is that ludwig was loading pre-trained models from huggingface hub at inference time when running in a container.

This also ensures that models saved by versions of Ludwig prior to #1877, or models which encountered an error during training will be restored without fetching a pretrained model from model hub.

[github-actions]

Unit Test Results

       6 files  ±0         6 suites  ±0   1h 57m 50s ⏱️ - 14m 34s
2 821 tests +1  2 787 ✔️ +1    34 💤 ±0  0 ❌ ±0 
8 463 runs  +3  8 357 ✔️ +3  106 💤 ±0  0 ❌ ±0 

Results for commit 4e5dfeb. ± Comparison against base commit ae25cc4.