[bugfix] support negative-only pointwise reranker data and fix evaluation metrics

[yinpu]

PR type

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

PR information

This PR fixes several reranker evaluation and data handling issues around pointwise training.

Background

The current reranker pipeline has a few gaps in pointwise evaluation/training:

• negative-only pointwise samples are dropped by the data collator, although they are valid
  for BCE-style pointwise reranker training
• evaluation metrics can regress when query group sizes are not inferred correctly
• metric gathering/publishing is not stable when group_sizes needs to be carried through
  evaluation

These issues can lead to incomplete training data usage and incorrect reranker metrics during
evaluation.

Changes

This PR includes the following fixes:

• support negative-only samples in the pointwise reranker data collator
• preserve group_sizes for pointwise reranker evaluation so query boundaries are explicit
• fix reranker evaluation behavior when query groups have variable sizes
• fix reranker metric gather logic to correctly handle (labels, group_sizes) tuples
• fix reranker metric publishing in evaluation loops while preserving the metric key prefix
• add regression tests covering collator behavior, metric calculation, and trainer evaluation
  flow

Affected areas

• swift/template/base.py
• swift/metrics/reranker.py
• swift/trainers/reranker_trainer.py

Test coverage

Added/updated tests:

• tests/train/test_reranker_collator.py
• tests/train/test_reranker_metrics.py
• tests/train/test_reranker_trainer.py

Covered cases include:

• pointwise reranker supports negative-only samples
• pointwise reranker supports positive-only samples
• listwise reranker behavior remains unchanged for negative-only data
• group_sizes is only emitted when needed
• metric calculation works with explicit query group boundaries
• trainer evaluation keeps metric_key_prefix
• gather logic preserves tuple labels for reranker metrics

Compatibility / Risk

This PR only changes reranker-specific collation and evaluation paths.

Main risk:

• unintended behavior change for existing reranker evaluation flows

Mitigation:

• regression tests were added for the affected pointwise/listwise collator and trainer paths

Experiment results

Unit tests added for reranker collator, metrics, and trainer regression cases.

[gemini-code-assist]

Code Review

This pull request introduces support for pointwise reranking and improves query boundary tracking using a new group_sizes attribute. Key changes include updating the reranker data collator to handle negative-only samples for pointwise loss, enhancing RerankerMetrics to calculate classification metrics (accuracy, precision, recall, and F1), and modifying the trainer to propagate group_sizes through the evaluation loop. Additionally, new test suites were added to verify the collator, metrics, and trainer logic. I have no feedback to provide.