Support slimmable (SlimmableContainer) files in bench_a2_fast

[jfsantos]

What

bench_a2_fast previously only accepted plain WaveNet .nam files and bailed on slimmable files. A2 models are shipped as SlimmableContainer models that bundle multiple standalone submodels (e.g. A2-Full at 8 channels and A2-Lite at 3 channels), so they couldn't be benchmarked with this tool.

This adds a --slim V (0.0–1.0) option that resolves a SlimmableContainer to a single submodel using the same rule as ContainerModel::SetSlimmableSize (first submodel whose max_value exceeds V, else the last), then runs the existing fast-vs-generic comparison on that submodel's config + weights.

Notes

• Plain WaveNet files are unaffected — passing --slim there just prints a note and benches normally.
• Omitting --slim on a container defaults to full size (last submodel).
• No library changes: each container submodel is already a standalone a2-shaped WaveNet spec, so it flows through the existing fast/generic paths directly.

Testing

Ran against a real A2 SlimmableContainer file:

• --slim 1.0 → submodel 1, A2 standard (8 ch), ~1.6× fast-vs-generic speedup
• --slim 0.2 → submodel 0, A2 nano (3 ch), ~6.5× fast-vs-generic speedup