Fix retry queue to re-run failed tests on Buildkite rebuild

[ianks]

Situation

When Buildkite rebuilds a job (new BUILDKITE_BUILD_ID), workers start fresh in a new Redis namespace — so their per-worker logs are empty. retry_queue builds its test list by intersecting the per-worker log with error-reports, and with an empty log that intersection is always empty. Workers would print "The retry queue does not contain any failure, we'll process the main queue instead", find the main queue exhausted too, and exit 0. The summary step would then read the still-populated error-reports and fail indefinitely.

The same problem affects automatic LSO retries (BUILDKITE_RETRY_TYPE=automatic), already fixed in #390, which relied on the same intersection.

Execution

Worker#retry_queue (worker.rb): When the per-worker log intersection yields nothing, fall back to redis.hkeys(key('error-reports')) directly. Any worker in the retry batch can now pick up failures from any other worker.

Retry#acknowledge (retry.rb): BuildRecord#record_success runs inside a redis.multi block and destructures the result array by position. Static#acknowledge was returning a plain Ruby value instead of queuing a Redis command, shifting all subsequent indices and breaking the stats delta correction. Now queues a SADD to processed so the positions stay correct.

reset_stats (build_record.rb): Added a purge of error-report-deltas entries belonging to the resetting worker. Without this, apply_error_report_delta_correction would subtract from already-zeroed counters on a same-worker retry, producing negative failure counts.

QueueEntry.format (queue_entry.rb): Now raises ArgumentError when file_path is nil or empty, and canonicalizes via File.expand_path. Previously, nil file paths were silently accepted and persisted into Redis — making entry keys non-reproducible across different working directories.

SingleExample#queue_entry and RSpec counterparts: Resolve file_path via source_location / example.file_path instead of hardcoding nil.