Chain primitives in Arroyo

[fpacifici]

Based on #132.

PR #132, introduced parallelism
for transformation operations via Arroyo multi processing in python.

There is a limitation in that model that prevents us from running
sequential transformation steps in the same process. Adding multiple
srequential Map operations and configuring them for multi processing
creates multiple process pools and run each step in parallel in separate
Arroyo strategies.

This PR expands the concept of segment allowing multiple consecutive
transformation steps to be chained together and be executed in the same
processes inside a multi process pool.
This should reduce the overhead and the memory implications of the
previous PR.

• Introduces TransformerChains as an abstraction to chain together steps.
• Wirte it up in the Rust Arroyo adapter.
• Add some documentation for pipeline configuration that explains the
  distribution model.