[RNE Rewrite] Use std::span instead of raw pointers in native C++ ops + tokenizer cosmetic refactor

[msluszniak]

Description

Applies the treatment from the FSMN-VAD framing op (#1298) to the remaining native ops that indexed tensor buffers through raw pointers: wrap the buffer in a std::span, take a subspan per row/lane, and use range-based algorithms instead of manual pointer arithmetic.

• math/operations.cpp — sigmoid/threshold become std::ranges::transform; softmax/argmax take a per-lane subspan instead of re-deriving offsets from a raw pointer on each access.
• cv/image_ops.cpp — channel-plane and colormap-pixel access go through subspans; the toChannelsFirst memcpy becomes std::ranges::copy.
• cv/box_ops.cpp — decodeToXyxy takes a std::span<const float, 4> row rather than four loose floats, removing the boxesPtr[idx * 4 + 0..3] unpacking at both call sites.

Byte-level spans are bounded by size_ (bytes) and element spans by numel_, so every subspan is bounds-checked against the real allocation.

cv::Mat constructions that take data_.get() are left as-is: they hand the whole buffer to OpenCV rather than doing pointer math, so a span adds nothing. Core files are untouched, per the issue.

Also drops the five repeated if (!result.ok()) blocks in nlp/tokenizer.cpp in favour of a local unwrap helper, mirroring the one in core/model.cpp (that one is file-local and typed to executorch::runtime::Result, so it cannot be reused for tokenizers::Result).

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

Refactor is behaviour-preserving, and verification so far is compile-only — not yet run on a device.

All four translation units compile clean under the project's strict warning set from .clangd (-Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion -Wold-style-cast -Wcast-align -Wshadow -Wdouble-promotion), with no new warnings:

cd packages/react-native-executorch
for f in cpp/extensions/{math/operations.cpp,cv/image_ops.cpp,cv/box_ops.cpp,nlp/tokenizer.cpp}; do
  clang++ -fsyntax-only $(tr '\n' ' ' < compile_flags.txt) -Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion "$f"
done

To exercise at runtime: run apps/computer-vision (classification/segmentation/detection cover softmax, argmax, applyColormap, resize, normalize, toChannelsFirst, and nms) and the tokenizer screen in apps/nlp.

Related issues

Closes #1308

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes