Bound the amount of data on the data plane by time.

This is especially important for Fn API reads, where each element represents a shard to read and may be very expensive, but many elements may be waiting in the Fn API buffer.

The need for this will be mitigated with full SDF support for liquid sharding over the Fn API, but not eliminated unless the runner can "unread" elements it has already sent. 

This is especially important in for dataflow jobs that start out small but then detect that they need more workers (e.g. due to the initial inputs being an SDF).

Imported from Jira [BEAM-5791](https://issues.apache.org/jira/browse/BEAM-5791). Original Jira may contain additional context.
Reported by: robertwb.

Start by tracing Fn API reads, the data-plane buffer, and how runners send elements to workers. Clarify the intended time bound and how completion should be verified, since the issue does not name files, tests, or a precise acceptance condition.

Written by the indexing model from the issue text.