Use python read for text dataset

[lhoestq]

As mentioned in #622 the pandas reader used for text dataset doesn't work properly when there are \r characters in the text file.

Instead I switched to pure python using open and read.
From my benchmark on a 100MB text file, it's the same speed as the previous pandas reader.

[thomwolf]

Ok :)

[thomwolf]

One thing though, could we try to read the files in parallel?

[lhoestq]

We could but I'm not sure this would help a lot since the bottleneck is the drive IO if the files are big enough.
It could make sense for very small files.

[lhoestq]

Looks like windows is not a big fan of this approach
I'm working on a fix

[thomwolf]

I remember issue #546 where this was kinda requested (but maybe IO would bottleneck). What do you think?

[lhoestq]

I think it's worth testing multiprocessing. It could also be something we add to our speed benchmarks

[BramVanroy]

I remember issue #546 where this was kinda requested (but maybe IO would bottleneck). What do you think?

It still would be interesting I think, especially in scenarios where IO is less of an issue (SSDs particularly) and where there are many smaller files. Wrapping this function in a pool.map is perhaps an easy thing to try.

[lhoestq]

Merging this one for now for the patch release