perf(continuations): Have bounded number of segment data while proving

[Nashtare]

Proving with continuations requires to first generate segment data for all segment payloads within a block. This has a negative effect on memory as these segments are persisted until consumed, which happens at a much slower rate.

We should consider having some buffered approach to not hold in memory more than N segment data ahead (where N most likely would be the runtime number of workers to make sure that they can all pick up a next job if they happened to finish their current one at the same time).

[atanmarko]

Related, may be solved by #627

[atanmarko]

In the PR #793, segment generation is performed separately from the proving, and new segments are communicated to the proving task with the channel. It is easy to limit their number by adding new semaphore to task generation of the number of segments in the batch that will be proven in parallel. I have decided not to do that (not to introduce new potential choking point), as we have already implemented block_pool_size limitation that manages the number of blocks proven in parallel.