fix(mnn): use getNumpyData() instead of getData() for output extraction (~3x rec speedup)

[todorangrg]

Problem

MNNInferSession.__call__ in rapidocr/inference_engine/mnn/main.py extracts
the session output via:

out_tensor = MNN.Tensor(out_shape, MNN.Halide_Type_Float, MNN.Tensor_DimensionType_Caffe)
output.copyToHostTensor(out_tensor)
return np.array(out_tensor.getData()).reshape(out_shape)

getData() marshals the tensor contents through a slow per-element Python
path rather than exposing the buffer directly. For a PP-OCRv6 medium
recognizer output of shape (1, ~230, 18710) (~17MB), this step alone took
~450ms on our test hardware (RK3588, OpenCL backend), compared to
~210ms for the actual GPU inference (runSession) that produced it — i.e.
result extraction was costing more than twice as much as inference itself.

On a real-world image with 39 detected text boxes, this pushed total
recognition time from an expected ~9s to ~22s.

Fix

Use the buffer-protocol-based getNumpyData() instead, which is already
exposed on the MNN Python binding:

result = out_tensor.getNumpyData()
result = np.array(result, copy=True).reshape(out_shape)

The explicit copy=True is required — getNumpyData() returns a view into
memory owned by out_tensor. Without an explicit copy, we observed
intermittent segfaults once out_tensor went out of scope / was reused
across calls with different session shapes. With the copy forced, this
was fully stable across repeated runs and varying input shapes/batches.

Results (RK3588, OpenCL backend, PP-OCRv6 medium recognizer)

	before	after
per-crop output extraction	~454ms	~10ms
per-crop total (__call__)	~690ms	~240ms
39-box real image, total rec time	~22s	~8.4s

The remaining ~10ms for extraction is a legitimate memcpy+reshape cost for
~17MB of data and is not further addressed here.

Testing

• Verified numerical output is identical to the old getData() path via
  np.allclose on sampled outputs.
• Verified stability across varying crop widths/batch sizes (no segfaults
  over repeated runs, including shape changes between calls).
• Benchmarked before/after on a real multi-box image end-to-end.

Notes

This affects every user of the MNN engine backend, independent of model,
backend device (CPU/OpenCL/Vulkan), or hardware — it's a fixed cost in the
Python-level extraction path, not specific to our setup.

[Copilot]

Pull request overview

This PR optimizes MNN backend output extraction in MNNInferSession.__call__ by switching from per-element getData() to buffer-based getNumpyData(), reducing Python overhead during post-inference tensor conversion.

Changes:

• Replace np.array(out_tensor.getData()) with np.array(out_tensor.getNumpyData(), copy=True) to accelerate output extraction.
• Keep an explicit copy during conversion to avoid lifetime-related crashes when out_tensor goes out of scope or is reused.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.