Accelerate FastAI Preprocessing + Fix TabularNN time_limit

[Innixma]

Issue #, if available:

Description of changes:

• Accelerate FastAI Preprocessing
• For datasets with many columns (>1000), this speeds up FastAI preprocessing by >300x when using pandas==1.5.3
• While the speedup is most large using pandas==1.5.3 due to pandas having major slowdowns in several functions, it is also major using older pandas versions that are not as slow (instead of 300x faster, it is 30x faster).
• This speedup can result in training time going from 1000 seconds -> 25 seconds, and inference time from 70 seconds -> 0.13 seconds. (robert dataset)
• This PR additionally fixes another major bug: TabularNN time estimates were off by a factor of 128x, causing many situations where TabularNN will skip training due to thinking it will run out of time when time_limit is specified. This bug was introduced in [Post 0.6][Tabular] make tabular nn dataset iterable #2395 and effects releases v0.6.1 and v0.6.2. The reason is because len(train_dataloader) switched from being train_rows/batch_size to train_rows.

TODO:

• Benchmark
• Create minimal reproducible example using only pandas and numpy, submit performance regression issue to pandas

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Job PR-2909-31ad1e1 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2909/31ad1e1/index.html

[liangfu]

Looks awesome!

FastAI model was the slowest model at inference time, and now it's catching up with other models, especially for large batch inference.

[github-actions]

Job PR-2909-2364806 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2909/2364806/index.html

[Innixma]

Benchmark result looks good! Pareto dominates mainline (2022_02_12 is mainline, 2022_02_14 is this PR, predict axis is based on batch_size=1):