Preferred aggregation temporality is under-specified

[jmacd]

What are you trying to achieve?

Whether through Views or another mechanism, the MetricReader needs a way to dictate the aggregation temporality that is used by default. For some exporters (OTLP, Console, In-Memory), the choice is something the user might care about, particularly for OTLP where there are memory costs associated with this decision.

What did you expect to see?

Some way that a user can control temporality, if they care. The current specification does not give strict requirements for "preferred aggregation temporality". One solution was outlined in #2314, where the View allowed to specify a strategy which was either "preferred", "cumulative" or "delta" (hypothetically). Another solution that was discussed in that PR would to let the temporality decision be defaulted on a per-instrument basis.

This applies to most of the exporters, makes the decision a "5-way function" for the five instruments that define the concept of aggregation temporality (i.e., the six instruments except Gauge).

This PR defines the requirements, but does not go as far as defining environment variable syntax for configuring the temporality on a per-instrument basis.