Parallelism across entire preprocessing pipeline

Write now we use multiprocessing for semantic mapping, but we could also just do it over the entire pipeline. Maybe sklearn has a built-in function for this?

Start by tracing the preprocessing pipeline from its current multiprocessing implementation for semantic mapping, then review how sklearn could support parallel execution across the remaining stages. The issue names no files or tests, so first identify the relevant entry points and establish how correctness and performance should be measured before deciding what “entire pipeline” parallelism means.

Written by the indexing model from the issue text.