Confirm that HTTP instrumentations can provide sampling-relevant attributes at creation time

[lmolkova]

In #1916 we require (MUST) sampling relevant attributes to be provided at span creation time:

https://github.com/open-telemetry/opentelemetry-specification/blame/f34abb07ee8b8e102ab4693d9d6cef1aad95fb2b/specification/trace/semantic_conventions/http.md#L94

We do it under assumptions:

• it's feasible
• reasonably efficient (assuming Sampling decison is too late to gain much performance #620 is fixed)

If there are more than a few (1-2?) instrumentations where these assumptions are wrong, we will change MUST to SHOULD for sampling-relevant attributes.

This issue is a blocker for HTTP semantic conventions stability.

[jmacd]

Assuming we can identify implementations that find the required attributes infeasible or inefficient, I think there is still a path forward as long as we can find an alternative in each language that is reasonably efficient.

It comes down to this: Reasonably efficient is necessary and feasible is sufficient. We only have to change MUST to SHOULD if there is no reasonably efficient implementation, so I think this boils down to #620.

In an optimistic sort of way, I think we should just solve #620 quickly. Lazy-evaluation is a widespread technique practically everywhere, and all we are doing is loosening the specification to allow languages to do anything natural to defer the cost of attributes so that it depends on whether the sampler cares or not.

[bogdandrutu]

Per the SIG meeting discussion, assign this to @lmolkova

[lmolkova]

This can be closed, we changed MUST to SHOULD in open-telemetry/semantic-conventions#422