Convert persist_source to a regular non-async operator

[antiguru]

Convert the persist_source decode_and_mfp operator to a plain Timely operator without using async. This avoids maintaining Timely state across yields and seems simpler to reason about in general.

This also takes the intended changes from #17156 to compact the output in-place using the ConsolidateBuffer. This way, we don't have to worry about large allocations or spending significant time in consolidating updates.

Checklist

• This PR has adequate test coverage / QA involvement has been duly considered.
• This PR evolves an existing $T ⇔ Proto$T mapping (possibly in a backwards-incompatible way) and therefore is tagged with a T-proto label.
• If this PR will require changes to cloud orchestration, there is a
  companion cloud PR to account for those changes that is tagged with
  the release-blocker label (example).
• This PR includes the following user-facing behavior changes:

[danhhz]

I'm supportive of the overall direction, but the original behavior of yield_fn was a suggestion from @frankmcsherry, so I'd love to get signoff from him on the change in behavior before looking at the rest of this too closely

[antiguru]

I can bring back the yield_fn if needed, there's no fundamental reason that we can't have it in this framework. The reason I removed it is because we never yield on time, only on number of records.

[vmarcos]

I will defer to storage folks regarding the details of the operator code; at least upon a comparison of the old and new versions, it seems to me that the logic has been kept unharmed.

Regarding the changes to ConsolidateBuffer, if we now introduce the function give_at, would it make sense to have a variation of new that does not require the output port to be specified, with an explicit Option being passed in? I am a bit worried of misuses where one gives a port in a multi-output operator, but calls give_at by mistake. With some extra checking, the implementation could potentially catch these problems?

[antiguru]

I updated the PR to use the yield_fn which was there previously. The PR is now best viewed with whitespace hidden!

[danhhz]

persist_source change LGTM! I didn't look at the buffer.rs change. sorry this has been such a saga!

[teskje]

LGTM

[vmarcos]

Took another look at give_at; it was actually fine, so this LGTM!