Improve parallel reader to minimize number of connections #124

What is the problem the feature request solves?

The current parallel reader splits a consecutive range into smaller pieces and reads them all in parallel. This method is fast but expensive as it uses a large number of connections which are not reused. On cloud storage this can increase the cost quite a bit.
Connections are not reused to avoid large seeks within a file. On cloud storage platforms like S3, long range seeks will automatically create a new http connection resulting in the same extra costs.

Describe the potential solution

A solution would be to use a set of reusable connections that read the consecutive ranges in parallel with no splitting of ranges. This approach has a disadvantage in that if there is a single large range then that range will be read by a single thread and will become a performance bottleneck. To overcome this, we can introduce a method to handle skew where such oversized ranges are split into smaller ranges.

Additional context

No response