Fix parallel test flakes

[jlamypoirier]

Summary

• allocate per-worker port blocks and dedicated streaming ports to avoid xdist port collisions
• use Hugging Face Hub HTTP backoff for Croissant metadata fetches while preserving live Hub test coverage
• improve the Croissant metadata assertion/log messages so endpoint/network failures are clear

Testing

• FAST_LLM_TEST_RESULTS_PATH=/tmp/fast_llm_tests/Fast-LLM" /Users/joel.lamy-poirier/Projects/Fast-LLM/venv/bin/python -m pytest -v -n 4 tests/data/test_streaming.py::test_streaming_dataset tests/data/test_streaming.py::test_streaming_sampled_dataset tests/data/test_streaming.py::test_data_streaming
• FAST_LLM_TEST_RESULTS_PATH=/tmp/fast_llm_tests/Fast-LLM" /Users/joel.lamy-poirier/Projects/Fast-LLM/venv/bin/python -m pytest -v -n 4 tests/data/test_streaming.py::test_data_streaming tests/data/test_streaming.py::test_run_data_streaming_distributed tests/data/test_streaming.py::test_data_streaming_distributed
• FAST_LLM_TEST_RESULTS_PATH=/tmp/fast_llm_tests/Fast-LLM" /Users/joel.lamy-poirier/Projects/Fast-LLM/venv/bin/python -m pytest -v -n 4 tests/data/test_preparator.py::test_dataset_preparator_from_hub
• FAST_LLM_TEST_RESULTS_PATH=/tmp/fast_llm_tests/Fast-LLM" /Users/joel.lamy-poirier/Projects/Fast-LLM/venv/bin/python -m pytest -v -n 4 tests/models/test_streaming.py::test_run_model_distributed_streaming tests/models/test_streaming.py::test_model_distributed_streaming (skipped locally: CUDA unavailable)
• git diff --check