[Refactor] Restructure test suite to match source layout and isolate benchmarks

[DhiraPT]

Summary

This PR refactors the testing infrastructure to separate correctness tests from performance benchmarks and establishes a strict directory mirroring strategy for better maintainability.

Key Changes

• Separation of Concerns: Moved performance/profiling code from tests/ into a dedicated benchmark/ directory (e.g., bench_index.py, bench_store.py).
• Infrastructure: Added conftest.py with pytest fixtures (e.g., cuda_device) and updated pyproject.toml with markers (@cuda, @distributed, @e2e).
• Strict Mirroring: Reorganized tests/ to exactly mirror the python/minisgl/ source structure, documented in the new tests/README.md.
• Coverage: Added unit tests for message/backend and message/utils (replacing the old test_serialize.py).