[Torchscript] Parallelized Text/Sequence Preprocessing

[geoffreyangus]

This PR parallelizes sequence tokenization for sequence and text features. This gives us better or equal inference throughput relative to vanilla Ludwig Model preprocessing.

One thing to note: there were/are strange interactions with torch.no_grad and the added parallelism. It seems that torch.no_grad affects some global flag that activates/deactivates gradient computation (link), and it seems that this flag is not properly reset after some scripted, parallelized operation.

The workaround has to do with the insight that gradients are not computed during preprocessing time and that the extraneous torch.no_grad statements could be removed (particularly around preprocessing and postprocessing). Applying the torch.no_grad context exclusively at the predictor stage of inference is enough to ensure that the module output tensors contain no gradients. Added tests confirm this. That said, we should keep this issue in mind if we decide to introduce parallelism in our ECD architecture.

[github-actions]

Unit Test Results

       6 files  ±  0         6 suites  ±0   2h 11m 29s ⏱️ - 19m 34s
2 886 tests +  6  2 840 ✔️ +  6    46 💤 ±0  0 ❌ ±0 
8 658 runs  +18  8 516 ✔️ +18  142 💤 ±0  0 ❌ ±0 

Results for commit 76310eb. ± Comparison against base commit 5193e0d.

♻️ This comment has been updated with latest results.