Pipeline physical resource limits

[jmacd]

This issue covers our interest in physical mechanisms for limiting various amounts of quantity in the pipeline, including concurrent request limits, item count limits, and memory space limits.

The issue #919 covers our interest in unifying the configuration plan for limits, i.e., and/or resource limits, topics like how we might configure either sort of limit based on tenant ID and/or signal type.

As the Phase 1 OTel-Arrow Collector components describe, https://github.com/open-telemetry/opentelemetry-collector/issues/9591, as investigated throughout https://github.com/open-telemetry/opentelemetry-collector/issues/12603, we can limit requests in certain ways through gRPC middleware, and in certain ways that we only know after applying the IPC reader. As a unifying principal, these space-based resource limits sometimes require a way to block the producer, and sometimes (usually after allowing some degree of blocking) to return an immediate failure.

This issue will link similar developments for request limits and admission controls generally.

Connected with #1244: we can free a resource reservation for memory as soon as the RETURN_DATA is dropped.