Benchmark baseline versus fused preprocessing on Android

[triasha72]

Summary

• benchmark the existing two-pass baseline against the single-pass fused FP32 preprocessing path on Android
• use a deterministic 262,144-element buffer to avoid timer-dominated tiny-input measurements
• report baseline/fused means, speedup, preprocessing drift, and downstream output drift through JNI/logcat
• extend repeated-device CSV capture with every comparison field
• add large-buffer baseline/fused equivalence coverage
• bump Android to version 0.1.5 (code 6)

Measurement boundaries

Preprocessing is measured separately from the fused preprocess-plus-reference-inference cold/warm timings. This remains the deterministic reference backend, not QNN/NPU evidence, and power remains unmeasured.

Verification

Bash syntax and diff checks pass locally. CMake is unavailable on the Mac, so GitHub CI must compile/test native and Android targets and pass 16 KiB alignment checks before merge.

[triasha72]

Superseded by the later merged end-to-end Android and release-acceptance PRs; the baseline-versus-fused preprocessing functionality is already present in main.