feat: add predictive autoscaling laboratory

[hadronzoo]

This draft adds an unpublished predictive autoscaling core, simulator, and WebAssembly adapter.

The core uses bounded Bayesian state, columnar scratch storage, and explicit SIMD operations. Criterion benchmarks cover realistic controller work.

The simulator separates regime inputs from the model. It models Kafka placement, fixed key concurrency, shared dependency limits, launch delay, and KIP-848 rebalance pauses.

Regime reports use vector figures and embedded posterior views. Generated reports remain outside Git through the ignored crates/scale-sim/reports/ directory.

The adversarial audit is still in progress. Current claim checks reject unsupported behavior for short bursts, seasonal demand, timer forecasts, and post-knee capacity evidence.