feat(recognition): add configurable confidence aggregation methods

[sneakybatman]

Summary

This PR adds support for configurable word-level confidence score aggregation methods in text recognition models. Previously, models used either arithmetic mean or minimum for aggregating character-level confidence scores into word-level confidence, with no way for users to customize this behavior.

Motivation

Different use cases may require different confidence aggregation strategies:

• Arithmetic mean: Good general-purpose default, balances all character confidences
• Geometric mean: More sensitive to low confidence characters, useful when any low confidence should significantly impact the word score
• Harmonic mean: Even more conservative, heavily penalizes low confidence characters
• Minimum: Most conservative approach, word confidence equals weakest character (good for high-precision requirements)
• Maximum: Most optimistic, useful when you want the best-case confidence
• Custom callable: Full flexibility for specialized use cases

Changes

• Add aggregate_confidence() utility function in core.py with support for 5 built-in methods plus custom callables
• Add ConfidenceAggregation type alias for type hints
• Add confidence_aggregation parameter to RecognitionPostProcessor base class
• Update all PyTorch PostProcessors: PARSeq, ViTSTR, CRNN, SAR, MASTER, VIPTR
• Update all TensorFlow PostProcessors: PARSeq, ViTSTR, SAR, MASTER
• Update remap_preds() for split crop handling to use configurable aggregation
• Add comprehensive unit tests (20 new test cases)

Usage Example

from doctr.models import recognition

# Use default aggregation (model-specific)
model = recognition.parseq(pretrained=True)

# Or customize at the PostProcessor level
from doctr.models.recognition.parseq.pytorch import PARSeqPostProcessor

# Use geometric mean for more conservative confidence scores
processor = PARSeqPostProcessor(vocab, confidence_aggregation="geometric_mean")

# Use custom aggregation function
import numpy as np
processor = PARSeqPostProcessor(vocab, confidence_aggregation=lambda probs: np.percentile(probs, 25))

Test plan

• All existing tests pass
• New unit tests for aggregate_confidence() function cover all 5 methods
• Tests verify correct handling of edge cases (empty arrays, single values, zeros)
• Tests verify custom callable support
• PyTorch postprocessor tests updated and passing
• TensorFlow postprocessor tests updated and passing