S3 or other file IO backends

**Is your feature request related to a problem? Please describe.**
With increasingly large datasets, "conventional" file storage is sometimes not the ideal way to store them.
For example, the Common Crawls are distributed on S3 buckets.

**Describe the solution you'd like**
To be able to integrate such datasets with ir_datasets, it would be great to be able to swap the IO "backend".
So for example, instead of opening a file from location `/path/to/file`, we could open files from `s3.example.com/bucket/path/to/file`.

**Describe alternatives you've considered**
Mounting S3 as a file system can be done but that would probably slow down file access. Also this is not officially endorsed by S3 storage providers like Amazon.

**Additional context**
I think that Hadoop/Yarn/Beam all support multiple underlying protocols for file access.