PySpark test suite is not self-contained

Summary

The uv workspace is meant to give an identical, reproducible test environment,
but the PySpark test suite depends on machine env state in two ways:

1. Spark/JVM configuration leaks in from the shell. SPARK_HOME and a
   family of related variables are honored by PySpark's launcher and the JVM,
   so a system-wide Spark install or exported JVM flags change or break
   every test session.
2. The JVM itself is not part of the venv. The pyspark wheel bundles
   Spark's jars and launcher scripts but no Java runtime; the tests silently
   rely on whatever java the machine resolves.

Both are the same underlying problem: the uv venv does not fully own the test
runtime.