InnoDB performs dirty read of TRX_SYS page before crash recovery

InnoDB startup is discovering undo tablespaces in a dirty way. It is reading a possibly stale copy of the TRX_SYS page before processing any redo log records.

A sign of this problem is the buf_pool_invalidate() call in innobase_start_or_create_for_mysql(). It should not be needed at all, and instead, in recv_recovery_from_checkpoint_start() we should assert that the buffer pool is empty before starting to apply redo log records.