docs(explanation): the quantization process — weights, activations, calibration

[michalharakal]

What

Adds an explanation/ article documenting SKaiNET's quantization process end to end, and a nav
entry under Explanation.

The existing docs cover weight quantization well (TurboQuant KV-cache, quantized SIMD kernels),
but nothing explains activation post-training quantization (PTQ) with calibration — what an int8
NPU/accelerator target needs. This page fills that gap.

Contents

• The two regimes side by side (weight-only block quant vs activation PTQ) and when each applies.
• The int8 matmul shape — quantize → i8 matmul (i32) → requantize — and the symmetric per-tensor
  scales Sa / Sb / So.
• Calibration: tapping the fp32 eager run with a TensorOps matmul decorator (Kotlin
  interface delegation) to capture per-tensor activation ranges over representative inputs, with the
  real API and a sample scales file.
• The honest finding on why QuantizedMatmul can't carry the activation-int8 path (eager +
  weight-only; TensorOps.matmul is dtype-homogeneous — no i8×i8→i32, no round), and the
  resulting split: the framework calibrates, the target backend emits int8.
• End-to-end pipeline sketch and follow-ups (tape-level i8 matmul, per-channel scales, observer
  notifyOp wiring).

Notes

• Documentation only — no code changes.
• Both xref: targets (quantized-simd-kernels, turboquant-kv-compression) verified to exist;
  uses the component-level {framework_name} attribute like the other pages.
• The calibration example is written as a reusable framework pattern (a generic matmul tap), not
  tied to any downstream consumer.

Closes #746.

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-747 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.