Implement Stage 17 hidden layer growth

[aarambh-darshan]

Summary

• Implement real Stage 17 hidden-layer growth by allowing additional hidden layers before the output layer.
• Preserve existing bound answer readouts by feeding the output layer from flattened hidden activations.
• Generalize training, backpropagation, diagnostics, and store validation for deep network topologies.
• Add Stage 17 CLI integration coverage, deep persistence coverage, and focused core/learn layer-growth tests.
• Update README, ROADMAP, ARCHITECTURE, CHANGELOG, and regenerated BENCHMARKS for Stage 17 completion.

Implementation details

• Network forward pass now records per-layer hidden activations and output activations in a cache.
• Hidden readouts include layer index, neuron index, global hidden index, and neuron id so diagnostics and bound memory remain stable after depth growth.
• Layer growth inserts a new hidden layer before output, expands downstream input widths, and appends output columns without rewriting older fact columns.
• Backprop computes deltas and gradients across all hidden layers instead of assuming a two-layer topology.
• Trainer growth now chooses deepest available hidden capacity or grows a new layer when hidden memory is saturated.
• Persistence validation now accepts variable hidden depth while preserving the existing file format version.

Verification

• git diff --check
• cargo fmt --all -- --check
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo build --workspace --release
• cargo bench -p manas-benches -- --quick
• cargo bench -p manas-benches -- --write-markdown BENCHMARKS.md
• bash demo.sh