$fluent->errors stocks all error logs unlimitedly.

Current implementation stores error logs in `$fluent->errors` unlimitedly. It would cause OOM killer if a "td-agent" process were gone(by segmentation fault, oom killer, etc.)
I know, fluent-logger-perl users should clear `$fluent->errors` buffer after sending a message. But, this issue is easily happen in a long running process, and hard to detect.

I suggest to add the hard limit to `$fluent->errors`.