Coalescing range requests *across* columns

Right now, each column needs to do it's own IO separately. But if these reads together are smaller than block_size of the object store, we should coalesce them.

Advice from @westonpace

Search for create_scheduler_decoder. There is a spawn in there (this is the scheduler thread). After it calls schedule_ranges or schedule_take you can call coalesce_outstanding_io (or something like that) which can be a new method you add to the EncodingsIo trait.
In the scheduling thread you have access to the EncodingsIo as config.io
Also, you might need to bump up fragment parallelism if you have hundreds / thousands of tiny fragments. The default is pretty high though (2 * io_parallelism)

Start by searching for create_scheduler_decoder and reading the scheduler thread around schedule_ranges and schedule_take. Inspect the EncodingsIo trait and config.io; done means outstanding reads across columns are coalesced when their combined size is smaller than the object store's block_size, with fragment parallelism considered for many tiny fragments.