Fix handling of empty metric collection cycles

[ocelotl]

Fixes #3198

[aabmass]

I think we should discuss the issue more in the SIG.

This PR would be a breaking change of the API. If we wanted to go this route, I'd recommend changing here to not call self._receive_metrics if they are empty instead of making it Optional:

opentelemetry-python/opentelemetry-sdk/src/opentelemetry/sdk/metrics/_internal/export/__init__.py

Lines 325 to 328 in 9d3d0f8

	self._receive_metrics(
	self._collect(self, timeout_millis=timeout_millis),
	timeout_millis=timeout_millis,
	)

[ocelotl]

I think we should discuss the issue more in the SIG.

This PR would be a breaking change of the API. If we wanted to go this route, I'd recommend changing here to not call self._receive_metrics if they are empty instead of making it Optional:

opentelemetry-python/opentelemetry-sdk/src/opentelemetry/sdk/metrics/_internal/export/__init__.py

Lines 325 to 328 in 9d3d0f8

	self._receive_metrics(
	self._collect(self, timeout_millis=timeout_millis),
	timeout_millis=timeout_millis,
	)

Right, fixed.

[ocelotl]

I am investigating #3277, marking this as a draft meanwhile

[aabmass]

While this might help for deltas, the SDK should be returning metrics every single collection for readers that are CUMULATIVE (the default). However that seems to not be working corrctly #3277 (comment)

Yes, noticed that now that I added another test case. I think this is a problem for the explicit bucket histogram aggregation who is resetting to zero all bucket counts after every collection even in the aggregation temporality is cumulative. I say we review this PR and address that issue separately.