Clarify A2 feature-test fixture

[sdatkinson]

Summary

• rename the misleading wavenet_a2_max.nam fixture to wavenet_a2_feature_test.nam
• update the fixture generator and tests to use the new name
• clarify that this generic feature-coverage fixture is not the eight-channel A2-Full architecture

This prevents performance results from the generic feature-stress model being mistaken for measurements of the specialized A2-Full path.

Verification

• Release run_tests passes
• touched C++ tests pass clang-format 19 validation
• actual A2-Full candidate performance was 0.19% faster than v0.5.4 by median across 30 interleaved runs
• the current A2-Full specialized path measured 1.27x faster than the generic WaveNet path at 64-sample blocks