Optimization for batched processing

[HYLcool]

• Extract the batched processing loop to a new function (e.g. compute_stats_batched), and decide the function branch to be called when processing data.
• Pros:
  • No need to modify each OP to support batched processing. Only need to set the arg _batched_op to True for each OP.
  • Developers can add new OPs in a single-sample-dict way, which is more friendly.
• Cons (for now):
  • Slightly slower data processing: 4 scenarios were tested and the results are shown below:

Scenario	OPs	Before	After
Simple Filter	text_length_filter	4.850s	5.340s (+10.10%)
Simple Mapper	whitespace_normalization_mapper	12.145s	12.575s (+3.54%)
With context (OP fusion)	word_num_filter & word_repetition_filter	80.456s	87.315s (+8.53%)
With rank (Model-based OPs)	image_text_similarity_filter	562.586s	590.502s (+4.96%)

[drcege]

Please sync the main branch.

[drcege]

LGTM