Bound per-cycle binlog purge to avoid LOCK_index write stalls

[vladislav-manafov-aiven]

Reducing binlog_retention_period on a write-heavy service made the next 60s purge tick collect the entire newly-eligible backlog and issue one PURGE BINARY LOGS over all of it. MySQL runs that under LOCK_index, which every committing transaction also needs, so a large backlog stalls all writes for the full purge duration (hours / >1 TB).

Cap each purge cycle by file count and total bytes so a large backlog drains as many small purges across successive ticks instead of one long stall. The cap trims the tail of the already-computed contiguous oldest-first prefix, so all existing safety semantics (stream safety, min_binlog_age_before_purge, replication state, no-GTID handling) are preserved. At least one file is always purged to guarantee progress.

Also emit myhoard.binlog.purged_per_cycle.{files,bytes} so the cap can be verified.

[alexole]

Please create a follow up to reduce interval when throttled to 1 second or even less, otherwise the rate of binlogs generation might be higher then removing and there is a risk of filling up the disk.