High performance direct runner

In documentation (https://beam.apache.org/documentation/runners/direct/) it is written that direct runner does not try to run efficiently, but it serves mostly for development and debugging.

I would suggest that there should be also an efficient direct runner. If Beam tries to be an unified programming model, for some smaller tasks I would love to implement them in Beam, just to keep the code in the same model, but it would be OK to run it as a normal smaller program (maybe inside one Docker container), without any distribution across multiple machines. In the future, if usage grows, I could then replace underlying runner with something distributed.

Imported from Jira [BEAM-2026](https://issues.apache.org/jira/browse/BEAM-2026). Original Jira may contain additional context.
Reported by: mitar.

Start with the Direct Runner documentation at documentation/runners/direct/ and review the linked Jira issue BEAM-2026 for the missing context. Clarify the intended efficient local execution model, including whether a single Docker container is in scope, before identifying the implementation entry point and acceptance criteria.

Written by the indexing model from the issue text.