Run premerge tests in 12 parallel paths [fast-ut] [reduced-it]

[res-life]

What changed

• split premerge into six independent unit-test paths and six independent integration-test paths
• shard each of the three logical integration-test configurations into two deterministic partitions
• add stable pytest sharding based on the full node ID using Java String.hashCode semantics
• apply reduced-IT each-choice selection before sharding and retain the existing fast-UT opt-in
• document and test the new shard controls

Why

The previous pipeline grouped multiple unit-test configurations and integration-test runs into three long-running paths. This serialized substantial work inside each path and limited premerge throughput. Separating the six unit-test executions from the six integration-test shards allows all twelve paths to be scheduled independently.

Impact

The default premerge test coverage is preserved while reducing the critical path when enough GPU workers are available. [fast-ut] continues to enable the parallel unit-test runner, and [reduced-it] continues to use deterministic each-choice parameter selection before the selected cases are divided between shards.

Validation

• python3 -m pytest -q -p no:spark_init_internal integration_tests/src/main/python/test_shard_selection_test.py integration_tests/src/main/python/reduced_it_selection_test.py
• bash -n jenkins/spark-premerge-build.sh
• python3 -m py_compile integration_tests/src/main/python/conftest.py integration_tests/src/main/python/test_shard_selection_test.py integration_tests/src/main/python/reduced_it_selection_test.py
• git diff --check

Full GPU premerge validation is delegated to this draft PR.