perf(ocr): reuse CTC argmax indices for confidence scores

[caitianxiao]

CTCLabelDecode scans the vocabulary twice: once for character indices and again for confidence scores. Reuse the argmax indices with np.take_along_axis to avoid the second scan. Character decoding and word-box handling are unchanged. This is separate from the logits-copy change in #5182.

Added regression tests for float16/32/64, array layouts, list inputs, ties, blanks, repeated characters, nonfinite scores and word boxes. python -m pytest -q tests/test_ctc_label_decode.py: 29 passed. Black 24.4.2, the repository-selected Flake8 rules and license-header checks also passed. The tests exercise the decoder classes without initializing inference backends.

On a local CPU microbenchmark with NumPy 2.3.2 and float32 logits shaped (16, 128, 8192), median conversion/index/score-extraction time dropped from 10.05 ms to 9.09 ms across 20 measured pairs. This measures postprocessing, not end-to-end OCR.

Microbenchmark

import statistics
import time
import numpy as np

x = np.random.default_rng(17).random((16, 128, 8192), dtype=np.float32)
samples = {False: [], True: []}
for i in range(24):
    for gather in ([False, True] if i % 2 == 0 else [True, False]):
        start = time.perf_counter()
        p = np.array(x)
        idx = p.argmax(axis=-1)
        scores = (np.take_along_axis(p, idx[..., None], axis=-1)[..., 0]
                  if gather else p.max(axis=-1))
        elapsed = time.perf_counter() - start
        if i >= 4:
            samples[gather].append(elapsed)
for gather, times in samples.items():
    print("gather" if gather else "max", statistics.median(times) * 1000, "ms")

[CLAassistant]

All committers have signed the CLA.