[Performance] Avoid redundant CTC logits copy during text recognition

[1443858742]

Summary

• replace the unconditional np.array conversion in CTCLabelDecode with
  np.asarray
• avoid copying NumPy logits already returned by the inference runners
• preserve conversion support for list and other array-like inputs
• add focused tests for zero-copy NumPy handling and array-like compatibility

Closes #5181.

Why

The Paddle static and HPI runner contracts return List[np.ndarray]. Calling
np.array(pred[0]) therefore allocates and copies the complete recognition
logits before CTC decoding. On a complex real-world page in our profiling,
19 recognition batches produced 794.45 MiB of cumulative logits.

Across 75 paired observations from 75 real document images:

Placement	Images	Official mean	np.asarray mean	Paired mean reduction	p95 before	p95 after
Local NUMA	41	1.1557 s	0.8484 s	25.47%	1.8380 s	1.3849 s
Cross-NUMA	34	3.1997 s	1.5222 s	31.26%	10.8966 s	2.4293 s

Direct conversion timing for the 794.45 MiB logits was 0.2472 s locally and
1.2089 s across NUMA with np.array, versus about 0.00002 s with
np.asarray.

All 75 paired outputs were identical in line count, character count, title and
full-text hashes, parsed OCR lines JSON, and downstream semantic fields JSON.

Compatibility

np.asarray returns the same object for an existing NumPy array and still
converts list and other array-like inputs. The decoder only reads preds
through argmax and max; it does not mutate the runner output.

Validation

• python -m pytest -q tests: 5 passed
• Black 24.4.2: passed
• Flake8 7.0.0 with the repository's selected rules: passed
• license header check: passed
• git diff --check: passed

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

lxy seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}