fix(omnivoice): remove erroneous transpose in OmniVoiceConv1d

OmniVoiceConv1d initialized weights as [out,k,in] (MLX format) but
then transposed them to [out,in,k] inside callAsFunction before
passing to MLX.conv1d. This permuted every conv1d weight in the
vocoder encoder/decoder, causing pure noise output even with
correctly-loaded checkpoint weights.

Remove the spurious transpose so the weight layout matches what
checkpoint and MLX.conv1d both expect.

Author: bestmrlee-max

Sources/MLXAudioTTS/Models/OmniVoice/OmniVoice.swift

@@ -1015,10 +1015,11 @@ final class OmniVoiceConv1d: Module {
     }
 
     func callAsFunction(_ x: MLXArray) -> MLXArray {
-        // Weight stored as [out, kernel, in] (MLX format); data flows in NCL [B, C, L]
-        // MLX.conv1d expects NLC [B, L, C] input and [out, k, in] weight
+        // Weight stored as [out, in, kernel] (PyTorch format) → transpose to [out, kernel, in] (MLX)
+        let w = weight.transposed(0, 2, 1)
+        // Data flows in NCL [B, C, L]; transpose to NLC for MLX conv1d, then back
         let xNLC = x.transposed(0, 2, 1)
-        var h = MLX.conv1d(xNLC, weight, stride: strideVal, padding: paddingVal)
+        var h = MLX.conv1d(xNLC, w, stride: strideVal, padding: paddingVal)
         if let b = bias {
             let n = b.size
             h = h + b.reshaped([n])