Process vs runtime metrics definition

[lmolkova]

Based on the SemConv WG on 7/8, separating actionable items from #1161 and #1035 (comment)

• We need to document that Process metrics need to be attached to the process resource.
  • the primary scenario for those metrics is an external observer (collector)
• Runtime instrumentations should not really report process metrics because:
  • in-proc instrumentations can't guarantee the resource attributes to be present
  • can't report pid because there could be duplication with resource attribute
  • runtime process metrics would duplicate ones coming from the collector.
  • one way for runtimes to report process metrics could be to have them as opt-in (resource concern remains).
• Runtime-specific conventions should define individual CPU metrics, even if they are very similar to process.cpu.time