Fixes prediction saving for models with Set output

[geoffreyangus]

This PR fixes prediction saving for Set features (#2199).

Prior to this change, the list of probabilities crossing threshold was converted to np.array with no dtype specified. If any of the probabilities crossed the threshold, the np.array generated would be the dtype of the probability, typically the dtype float32 commonly used by PyTorch. However, if NO probabilities crossed the threshold, the np.array generated would be dtype float64, which is the default for np.array. This PR ensures that the dtype of the generated numpy array is float32.

[github-actions]

Unit Test Results

       6 files  +    1         6 suites  +1   2h 35m 28s ⏱️ + 35m 17s
2 911 tests +    3  2 865 ✔️ +    3    46 💤 ±  0  0 ❌ ±0 
8 733 runs  +177  8 591 ✔️ +154  142 💤 +23  0 ❌ ±0 

Results for commit b4531af. ± Comparison against base commit 236cf54.

♻️ This comment has been updated with latest results.

[abidwael]

Thanks for putting this together Geoffrey! Left a comment about using text_feature, but otherwise should be good.