Promtail metric vector improvement.

[cyriltovena]

As of today we've by-passed a lot of the prometheus golang client to build our own metric engine. This allowed us to build metric on the fly for each stream created by promtail.

However in the long term, streams are volatile (new stream are created and disappear every time a file is rolling) which could cause some memory issues if we were to keep too many metrics for some streams.

Since those metrics are created in one place https://github.com/grafana/loki/blob/267dbe0085a7dbff8062b850f2e5776268acf4d2/pkg/logentry/metric/metricvec.go we should register them into a global map and run some sort of stale metric manager that removes metrics that haven't been updated for a while (1~5min, could be configurable)

However it should still work if the metrics comes back suddenly.

/cc @slim-bean @rfratto