feat: add opt-in per-layer residual diagnostics

[DivyamTalwar]

Summary

• expose realized beta, axis-gain, projection-error, and update-norm metrics
• collect separate attention/MLP residual reports for every layer
• keep instrumentation disabled by default to avoid accelerator synchronization overhead
• verify the beta=1 projection identity, output parity, bounds, and JSON serialization

Acceptance

• instrumented and uninstrumented logits agree
• beta stays in [0,2] and |1-beta| stays in [0,1]
• beta=1 reaches a constant scalar projection target within 1e-5
• CI must pass before merge