Revert A2 fast optimizations

[sdatkinson]

Summary

This PR reverts the NAM/wavenet/a2_fast.cpp changes from #277 and adds example_models/A2.nam as an example A2 model.

Context

Follow-up testing in #290 showed that the #277 A2 fast-path optimizations are highly compiler/target dependent. Some compiler contexts see speedups, while others regress significantly.

On Apple M1 + AppleClang 16, benchmarking example_models/A2.nam against the parent of #277 showed:

• slim=1.0 / 8ch: ~29% slower by mean, ~30% slower by median
• slim=0.0 / 3ch: ~2% faster by mean, ~3% faster by median

Given the 8-channel regression, this restores the prior a2_fast.cpp implementation while we decide whether to reintroduce these paths behind compiler/configuration flags.

Changes

• Revert only the A2Fast: portable conv + memory optimizations (up to 1.6x on channels=8) #277 changes to NAM/wavenet/a2_fast.cpp
• Add example_models/A2.nam
  • SlimmableContainer
  • 3-channel submodel at max_value=0.5
  • 8-channel submodel at max_value=1.0

Validation

Built benchmodel successfully with:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --target benchmodel

diff

% git diff 9c7b185 -- NAM/wavenet/a2_fast.cpp 
diff --git a/NAM/wavenet/a2_fast.cpp b/NAM/wavenet/a2_fast.cpp
index 8fd8fc1..67093c8 100644
--- a/NAM/wavenet/a2_fast.cpp
+++ b/NAM/wavenet/a2_fast.cpp
@@ -66,10 +66,10 @@ public:
   ~A2FastModel() override = default;
 
   void process(NAM_SAMPLE** input, NAM_SAMPLE** output, int num_frames) override;
+  int GetPrewarmSamples() override { return _prewarm_samples; }
 
 protected:
   void SetMaxBufferSize(int maxBufferSize) override;
-  int PrewarmSamples() override { return _prewarm_samples; }
 
 private:
   struct Layer