Fix: Don't skip saving the model if the save path already exists.

[justinxzhao]

Without this change, the model is only saved the first time save() is called (at the end of the first round of evaluation).

This is a bug that was introduced in #2027. This was not caught by any model loading unit tests like tests/integration_tests/test_model_save_and_load.py because the best saved weights are reloaded automatically at the end of train(), so there's never a discrepancy between the model returned from train() and models loaded from disk.

@jimthompson5802 will send a follow up PR that adds a test that catches this regression specifically.

Closes #2259

[jimthompson5802]

I confirmed that this fix addresses the issue I found.

[github-actions]

Unit Test Results

       6 files  +    1         6 suites  +1   2h 31m 2s ⏱️ + 30m 36s
2 930 tests +    1  2 884 ✔️ +    1    46 💤 ±  0  0 ❌ ±0 
8 790 runs  +174  8 648 ✔️ +151  142 💤 +23  0 ❌ ±0 

Results for commit c38aad9. ± Comparison against base commit 9c58d5e.

♻️ This comment has been updated with latest results.

[jimthompson5802]

@justinxzhao I added the test we talked about into tests/integration_tests/test_model_save_and_load.py::test_model_weights_match_training

I confirmed that this test will pick up the situation when the wrong weights are loaded from checkpoint.

Assuming we get a clean run in the CI test, this PR is good to go.