fix(infer): skip invalid dataset samples

[RerankerGuo]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Fixes #7045.

Background

InferArguments.strict defaults to False, but SwiftInfer did not
propagate that setting to its inference engine. Template encoding
therefore still raised on the first invalid image, audio, or other
sample instead of allowing the remaining dataset to continue.

There was a second failure path when strict mode was disabled manually:
the engine returned the per-sample exception in its aligned response
list, and _batch_infer() attempted to access resp.choices, raising
AttributeError.

Changes

• Propagate args.strict to the inference engine created by
  SwiftInfer.
• Skip exception responses while assembling dataset inference results,
  with a warning that preserves the original error message.
• Add CPU regression tests for strict-setting propagation and mixed
  valid/invalid batch results.

Verification

• PYTHONPATH=. .venv/bin/python -m unittest tests.infer.test_infer_pipeline
• PYTHONPATH=. .venv/bin/python tests/run.py --pattern test_infer_pipeline.py
• PYTHONPATH=. .venv/bin/python -m unittest tests.infer.test_infer_pipeline tests.general.test_data_preprocess.TestProviderMessagesPreprocess tests.general.test_data_preprocess.TestRejectedMessagesPreprocess
• .venv/bin/pre-commit run --all-files
• git diff --check

Impact

The change affects dataset inference when strict=False. Valid samples
retain their existing outputs and labels; invalid samples are logged and
omitted from the result list. Strict inference and deployment behavior
remain unchanged.

Experiment results

Before the fix, the regression tests failed because the engine had no
propagated strict value and because an exception response was treated
as a normal completion. After the fix, both cases pass and valid samples
before and after the failed entry are preserved.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.