[iohzrd]

Replaces the previous version of this PR (99a2871). That version sent the request from the template thread through a pending-request slot shared with the coinbaser thread's fetch. Two requests could be outstanding at once, so a reply could be matched to the wrong request, and an unanswered request blocked later fetches for the same template. This version keeps the coinbaser thread as the only sender and does not change datum_protocol_coinbaser_fetch.

On a new block or an urgent work update, update_stratum_job publishes a JOB_STATE_EMPTY_PLUS job, waits up to 4 s for its empty work to be sent, then publishes a JOB_STATE_FULL_PRIORITY_WAIT_COINBASER job from the same template. The coinbaser request was sent only after the second job was published, so its round trip to the pool did not overlap the empty work send. Miners mine on a coinbase that pays only the pool until the reply is applied.

• datum_coinbaser_prefetch: when the coinbaser thread sees an EMPTY_PLUS job, it fetches the split for that job's value and previous block into a static job. Skipped unless datum_protocol_is_active(), the same condition as a job's own fetch.
• datum_coinbaser_prefetch_take: the next FULL_PRIORITY_WAIT_COINBASER job copies the outputs and coinbaser id if its value, previous block and DATUM session generation match. Otherwise it sends its own request, as before. Each take clears the stored split, so at most one job uses it.
• datum_protocol_session_generation() returns the session counter, so a split fetched on one session is not used on the next.
• When the first request is answered, one request is sent per block, the same as before.
• Tests: the matching job takes the split once; another session, previous block, value or job state does not; a fetch with no outputs is not taken; a non-matching job clears it.

Unchanged: the reply carries only the value, so a late reply to an earlier request with the same value and a different previous block is accepted, as on master.

Merges without conflicts with #9, #13, #21 and #24 separately, and with #9, #13 and #21 together. -Wall -Werror build and --test pass on each. With #9, the fetch used here also catches a reply that arrives before the wait starts.

[iohzrd]

Closing. The prefetch saves at most the shorter of the empty work send time and the round trip to the pool, which is a few milliseconds per block. The case that costs the split is a coinbaser reply that arrives after the 5 s timeout: the reply is discarded and miners keep the pool-only coinbase until the next standard job. That needs a separate change.