Update Python SDK to construct Dataflow job requests from Beam runner API protos

Currently, portable runners are expected to do following when constructing a runner specific job.

SDK specific job graph -\> Beam runner API proto -\> Runner specific job request

Portable Spark and Flink follow this model.

Dataflow does following.

SDK specific job graph -\> Runner specific job request

Beam runner API proto -\> Upload to GCS -\> Download at workers

We should update Dataflow to follow the prior path which is expected to be followed by all portable runners.

This will simplify the cross-language transforms job construction logic for Dataflow.

We can probably start this by just implementing this for Python SDK for portions of pipeline received by expanding external transforms.

cc: [~lcwik] [~robertwb]

Imported from Jira [BEAM-10012](https://issues.apache.org/jira/browse/BEAM-10012). Original Jira may contain additional context.
Reported by: chamikara.