Resolve most skipped unittests

[HYLcool]

• unit test opt:
  • open most skipped unit tests
    • 20 unit tests for API-based OPs are all opened again by setting the API info with environment variables
    • 1 unit test for video_motion_score_raft_filter is opened again by setting more appropriate thresholds to handle different model outputs on different hardware, which is normal.
    • 7 unit tests for model-based OPs are opened again by decreasing the number of processors to avoid OOM of GPU memory.
  • increase the shm_size to 128G to avoid implicit OOM issue with exit code 137 (ref)
  • for now, there are still 9 unit tests that are skipped left to be resolved in future work
    • 2 unit tests are skipped due to randomness from system resource utilization (adapter & monitor)
    • 1 unit test is skipped due to an unsolved encoding problem (nlpcda_zh_mapper)
    • 1 unit test is skipped due to OOM of GPU (video_captioning_from_summarizer_mapper, which requires more than 50GB mem)
    • 5 unit tests are skipped due to an unsolved vllm error for distributed inference (similar to [Bug]: vllm.LLM does not seem to re-initialize for distributed inference with subsequent models with Offline Inference vllm-project/vllm#9727) (generate_*_mapper & optimize_*_mapper)
• others:
  • add default mem_required for some model-based OPs
  • update sampling_params for vllm-based OPs due to vllm and transformers having similar inference parameters with different names.
  • update device_map specification for latest diffusers library.
  • fix some bugs & typos

[yxdyc]

LGTM