FR: introduce metric_buffer_size_limit

When running in K8S, telegraf is often running with a pod memory limit. Telegraf configuration should be done such way, that it wont trigger OOM. Likely, one of the biggest memory consumer is metrics buffer keeping metrics accumulated before flushes.

Current tunable `metric_buffer_limit` configures number of metrics which can be stored in the buffer. It is not obvious how to use that to keep telegraf memory under desired limit to prevent OOM: each metric has variable number of labels which in turn consume variable amount of memory. Currently `metric_buffer_limit` needs to be set empirically and final value heavily depends on type of metrics collected, which can change over time.

If telegraf could estimate total memory usage of all metrics in the buffer, then introducing `metric_buffer_size_limit` tunable would allow better control over memory utilization.

Start by locating the implementation and tests for the existing metric_buffer_limit setting, then trace how buffered metrics are represented and sized. Define how metric_buffer_size_limit should estimate memory and how it interacts with the existing count limit; done means the behavior is documented and covered by tests for variable-sized metrics.

Written by the indexing model from the issue text.