Cache A2 prewarm state

[sdatkinson]

Summary

• cache one steady-state input sample for each A2 convolution and rebuild ring histories from it on later prewarms
• support both the specialized A2-fast implementation and the unconditioned generic WaveNet fallback
• preserve full silence processing for conditioned WaveNets, whose nested DSP state cannot be reconstructed from convolution caches alone
• use the correct A2-Lite and A2-Full terminology throughout the fast-path code, tests, and benchmark output

Tests

• cmake --build build --target run_tests -j4 && ./build/tools/run_tests
• full debug suite with NAM_ENABLE_A2_FAST=OFF
• full debug suite with NAM_A2_RING_MODE=0
• A2-Lite and A2-Full tests first prewarm through legacy silence processing, process more than a full receptive field of audio, restore the cached prewarm state without allocations, and require exactly identical output when processing the same audio again; both A2-fast and generic WaveNet paths are covered

Resolves #318

[sdatkinson]

Performance metrics on my M1:

Model	Old code: 100 resets	New code: 100 resets	Per reset, old → new	Speedup
A2-Lite	72.981 ms	0.859 ms	729.81 → 8.59 µs	85.0×
A2-Full	517.224 ms	1.123 ms	5,172.24 → 11.23 µs	460.6×