Refactor testing framework, add ring attention benchmark

[haok1402]

• Removed legacy testing scripts for the MLA operators, for which FA4 handles it now.
• Refactored the testing framework to launch torch.distributed via mp.spawn with pytest integration, so future operator tests can follow this pattern. Yet to be extended if unit tests require a multi-node setup.
• Revised the ring attention benchmark and test:
  • Prepares for the next PR, which will address workload balance for standard ring attention.
  • Benchmark captures an nsys profile of one iteration per scenario (bracketed by cudaProfilerStart/Stop).
  • Added a real-scenario benchmark case: Qwen3-30B-A3B GQA dims at 32K context with CP=4.

[gemini-code-assist]

Code Review

This pull request introduces a performance benchmarking suite and a structured testing framework for ring attention. Key additions include a Python benchmark script and a shell launcher for latency measurement and profiling, along with a new unit test suite that validates ring attention against a dense baseline. The PR also refactors distributed testing utilities and updates model import paths. Reviewer feedback focuses on improving robustness and accuracy, specifically by suggesting assertions for sequence length divisibility, optimizing CUDA event management in benchmarks, enhancing CLI argument validation, and applying Bash quoting best practices.