Make convert_predictions() backend-aware

[hungcs]

discovered this from the predibase side - predictions can be dd.Series, and have to be converted to pandas

    pred_df = model.predict(inputs_df, return_type=dict)[0]
  File "/Users/hw/mambaforge/envs/based/lib/python3.8/site-packages/ludwig/api.py", line 803, in predict
    converted_postproc_predictions = convert_predictions(
  File "/Users/hw/mambaforge/envs/based/lib/python3.8/site-packages/ludwig/data/postprocessing.py", line 108, in convert_predictions
    return convert_fn(
  File "/Users/hw/mambaforge/envs/based/lib/python3.8/site-packages/ludwig/data/postprocessing.py", line 127, in convert_to_dict
    values = np.stack(values.to_numpy())
AttributeError: 'Series' object has no attribute 'to_numpy'

[github-actions]

Unit Test Results

       6 files  +    1         6 suites  +1   2h 17m 22s ⏱️ + 37m 44s
2 880 tests ±    0  2 834 ✔️ ±    0    46 💤 ±  0  0 ❌ ±0 
8 640 runs  +161  8 498 ✔️ +138  142 💤 +23  0 ❌ ±0 

Results for commit 43c9050. ± Comparison against base commit e3c03c9.

♻️ This comment has been updated with latest results.