[Post 0.6][Tabular] make tabular nn dataset iterable

[liangfu]

Description of changes:
This PR converts map-based tabular nn dataset class into an iterable dataset, which makes batch-based data loading significantly faster.

For instance, loading 9769 rows in test.csv has been reduced from 49 ms to 9 ms.

Data loading time has been measured with following code snippet.

        tic = time.time()
        subtotal = 0
        for batch_idx, data_batch in enumerate(val_dataloader):
            tik = time.time()
            preds_batch = self.model.predict(data_batch)
            preds_dataset.append(preds_batch)
            subtotal += time.time() - tik
        total = time.time()-tic
        print(f"elapsed (dataloader): {(total-subtotal)*1000:.0f} ms")
        print(f"elapsed (predict): {subtotal*1000:.0f} ms")

Under Linux, we are able to achieve 40% overall time savings for batch_size>10000.

See following chart for more benchmark results.

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-2395-8f85dcb is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2395/8f85dcb/index.html

[liangfu]

cc @tonyhoo @Innixma

[Innixma]

Looks very good! I am able to reproduce the speedups on multiple datasets, nice work!!

[github-actions]

Job PR-2395-2341393 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2395/2341393/index.html

[Innixma]

Were we dropping last before this PR?

[liangfu]

This depends on the drop_last argument in the data loader.

[Innixma]

I think I found the answer, we were, if you look at deleted code lines

[liangfu]

We won't drop_last or shuffle when is_test==True, we were dropping last and shuffle the dataset for training. The behavior is maintained to be consistent.