Remove redundant dataset utilities

[krishnakalyan3]

I have also addressed your comments on torch.random #1022 (comment).
Fixes #1081

[mthrok]

Thanks for working on this but you need to get rid of all the internal use.
Please run tests by pytest -vs test/torchaudio_unittest/datasets/

[krishnakalyan3]

Hi, @mthrok looks like I am missing some context. Can you please let me know why unicode_csv_reader should be removed?.
If removed what should I be replacing it with?.

[mthrok]

These utilities were added (for internal use) to ease the difference between Python 2 and Python 3.

But now that Python 2 support is officially dropped, we do not need these utilities. Instead we can use the standard module as-is.
In Python 3 os.makedirs function has exist_ok argument, and csv module can handle unicode natively without the hack of changing the field_limit_size.

So we just need to replace the usages of these functions with standard module. Then we can remove the utility functions themselves.

[krishnakalyan3]

@mthrok can you please review this.

[mthrok]

Looks good. Thank you so much!

[krishnakalyan3]

@mthrok can you please recommend some other (easy - medium complexity) issues that I could work on?.

Cheers,
Krishna

[mthrok]

@krishnakalyan3 Sure. Thanks for offering :) Let me think of an issue. Meanwhile, if you can join PyTorch's slack channel, we can discuss collaboration there too. Communication will be faster that way. https://pytorch.org/resources/

[mthrok]

@krishnakalyan3

How about #910 (comment) ?
I consider this as not easy in terms of complexity because there are somewhat open-ended perspectives in the specification, so we need to iterate on the design.