[android] Re-benchmark wav2vec2 bench app after onnxruntime-android 1.22.0 -> 1.28.0 (PR #205) to capture KleidiAI ARM64 gains

[grzanka]

Context

Dependabot PR #205 bumps com.microsoft.onnxruntime:onnxruntime-android 1.22.0 → 1.28.0 in
bench/android/wav2vec2/app/build.gradle. That app (BenchActivity.java:3-5,95) uses only the
stable Java bindings (OrtEnvironment, OrtSession, default empty SessionOptions) with
abiFilters "arm64-v8a" — plain CPU EP, no NNAPI/XNNPACK configured.

Across the 1.23.0–1.28.0 release notes, ARM64 CPU inference got several KleidiAI kernel upgrades
that apply automatically with no app code changes:

• 1.23.0: enhanced SGEMM/IGEMM/dynamic-quantized-MatMul performance, esp. Conv2D on SME2 hardware
• 1.25.0: BF16 SME2 kernel integration, asymmetric 4-bit MatMulNBits on ARM64
• 1.28.0: new ARM64 2-bit weight CPU kernels

docs/android-asr-runtime-bench.md already has committed baseline numbers for this exact app on
onnxruntime-android 1.22.0. Since the model it loads (model_quantized.onnx, wav2vec2-base-960h) is
quantized, these are exactly the kernels that would benefit.

Ask

1. Merge PR build(deps): bump com.microsoft.onnxruntime:onnxruntime-android from 1.22.0 to 1.28.0 in /bench/android/wav2vec2 #205 (onnxruntime-android 1.22.0 → 1.28.0).
2. Rebuild bench/android/wav2vec2 and re-run the existing benchmark harness on the same
   device/test set used for the baseline numbers already in docs/android-asr-runtime-bench.md.
3. Update docs/android-asr-runtime-bench.md with the new numbers, noting whether the KleidiAI
   kernel upgrades above produced a measurable speedup (or regression) versus the 1.22.0 baseline.

This is a "capture the free win" task, not a new feature — the version bump alone may already
improve on-device latency for this benchmark without any code change.