Add Parquet loader + from_parquet and to_parquet

[lhoestq]

Continuation of #2247

I added a "parquet" dataset builder, as well as the methods Dataset.from_parquet and Dataset.to_parquet.
As usual, the data are converted to arrow in a batched way to avoid loading everything in memory.

[lhoestq]

pyarrow 1.0.0 doesn't support some types in parquet, we'll have to bump its minimum version.

Also I still need to add dummy data to test the parquet builder.

[lhoestq]

I had to bump the minimum pyarrow version to 3.0.0 to properly support parquet.

Everything is ready for review now :)
I reused pretty much the same tests we had for CSV

[albertvillanova]

Thanks @lhoestq.

Only some small comments/questions...

[albertvillanova]

Are we sure that it is a good idea to stop supporting pyarrow < 3.0.0? Just to be sure of this choice... Maybe a softer option would be to set this requirement only for the users that want to use Parquet files...

[lhoestq]

Good point, I just checked and there are still many projects that use pyarrow 2.0.0 and 1.0.1
I'll make the change

[albertvillanova]

Some little fixes.

[lhoestq]

Done !
Now we're still allowing pyarrow>=1.0.0, but when users want to use parquet features they're asked to update to pyarrow>=3.0.0