Solana performance improvement

[sistemd]

As a consequence of how our BlockStreamer trait works, Solana currently downloads many CAR files in parallel and then processes them in parallel. However, when it comes to downloading data, doing things in parallel is actually slightly slower than doing them sequentially because the limit is network bandwidth - parallelism does not help with speed. What would be faster is if there was only ever one download going on at any time, and the workers were only processing files in parallel rather than downloading them. This means that the download happens in parallel with processing, so we're never stuck doing "nothing" waiting for a file to download.

This should result in a neat performance gain, but it will probably require some pretty significant refactoring for the BlockStreamer trait and surrounding code.

[sistemd]

Another idea (due to @LNSD) is to split the range of all epochs and process them in parallel without ever writing them to disk. This works because we only ever need to read CAR files sequentially. For example, if there are 800 CAR files, then one worker will process files 0-100, the next 100-200, and so on. And each worker will never store to disk. (Alternatively we can use a queue/channel to distribute the epochs to the workers.)

However, it's not immediately clear which approach is faster without trying.