Fixed scalar ggml_vec_cvar_f32

[sandrohanea]

Hello @ggerganov, @taronaeo,

While upgrading from whisper.cpp v1.8.0 → v1.8.1, I noticed that the non-SIMD (scalar) code path started producing incorrect results in whisper.net CI.
The last successful run was just before commit d83fef35, which introduced the new ggml_vec_cvar_f32 optimization (llama.cpp PR #15953).

After this change, the scalar branch wrote the squared residuals into y[i] instead of the centered values.
This caused normalization in ggml_compute_forward_norm_f32 to divide (x − mean)² by √(var + eps) instead of (x − mean), producing invalid activations (resulting in -inf logits and !!!!!! transcriptions).

The SIMD branches (AVX/SSE/NEON/VXE) were unaffected because they correctly stored the centered values before squaring for the sum.

This PR restores the correct behavior for the scalar path:

for (; i < n; ++i) {
    float d = x[i] - mean;
    y[i] = d;
    sum += (ggml_float)(d * d);
}

Verified in this pipeline run — whisper.net now produces expected transcriptions again.

[ggerganov]

Yes, sorry about that. This was fixed recently in llama.cpp: ggml-org/llama.cpp#16558

Will sync it later tonight/tomorrow.

[sandrohanea]

Missed that fix, thank you!