[Feat] OP-wise Insight Mining

[HYLcool]

• Support OP-wise Insight Mining to check the significance of stats/meta changes before and after each OP with T-Test measure.
• It helps users to know how each OP influences their datasets.
• Add a new measure, RelatedTTestMeasure, to measure T-Test for two related distributions on their histograms with the same bins.
• It analyzes small batches of the original dataset and the current dataset after each OP processing. After all OPs are processed, it collects all the analysis results and mines the stats/meta change significance insights from them.

Core related:

• Analyzer: Support analysis on str tags/categories/stats in stats and meta fields; support passing an existing Dataset object to the run method of analyzer instead of loading from disk.
• Monitor: Remove error messages from the resource monitor process that are caused by the main process errors, which helps to make logs clearer.
• Exporter: export meta along with stats if export_stats is True.

Others

• Add two registries, Non-stats Filters and Tagging Operators, to decorate the Filters that don't produce any stats and OPs that produce tags or other infos in meta field. For now, they are:
  • Non-stats Filters: specified_field_filter, specified_numeric_field_filter, suffix_filter, video_tagging_from_frames_filter (produces tags instead of stats).
  • Tagging Operators: video_tagging_from_frames_filter, image_tagging_mapper, video_tagging_from_audio_mapper, video_tagging_from_frames_mapper

⚠️ Notice

This PR is based on the PR #512 . Please review that PR first before reviewing this one.

[yxdyc]

LGTM, plz see the minor comments.