Implements#12221

When stats from agents are collected, they are aggregated by 4 dimensions:agent_name,template_name,workspace_name, andusername. This can result in some very high cardinality metrics being scraped by Prometheus in large environments.

This PR adds the ability to tune which labels are include in this aggregation, therefore reducing the cardinality.

For example:

agent_sessions_total{agent_name="main",magic_type="ssh",pty="yes",template_name="docker",username="danny",workspace_name="test1"} 6agent_sessions_total{agent_name="main",magic_type="ssh",pty="yes",template_name="docker",username="danny",workspace_name="test2"} 5

With hundreds of active workspaces, each having a unique name, the cardinality may be unacceptable. In this case the operator may choose to configureCODER_PROMETHEUS_AGGREGATE_AGENT_STATS_BY=username to rather aggregate by user, summing all the metrics' values and only producing a single metric series:

agent_sessions_total{magic_type="ssh",pty="yes",username="danny"} 11

Multiple labels can be provided, e.g.CODER_PROMETHEUS_AGGREGATE_AGENT_STATS_BY=username,agent_name

agent_sessions_total{agent_name="main",magic_type="ssh",pty="yes",username="danny"} 11

The current behaviour remains the default; if no value is passed toCODER_PROMETHEUS_AGGREGATE_AGENT_STATS_BY there will be no change.

Note to reviewer: I made a sweeping refactor to define the label names in a single location instead of duplicating them, which may make the PR larger than it seems.