apertus: add 1.5 Omni model and media support

[pd95]

Motivation

Apertus 1.5 Omni models should import text and media tensors and support native
ordered image/audio conversations through MLX.

Dependency

Base: feature/gemma4-mlx-multi-media (#17).

Changes owned by this PR

• Add Apertus 1.5 runtime, import, parser, renderer, and thinking controls.
• Import media tokenizers and implement native ordered image/audio processing.
• Support 8B/70B configuration, MP3, memory limits, and capability gating.
• Retain staged dependencies, request state, and media history across runner
  sweeps.

v0.32.11 architecture

The implementation uses official lazy base.MediaModel behavior, model-owned
caches, media-aware prefill, and runner-owned evaluation. A small staged-media
contract preserves graph-cut dependencies; official Nemotron routing remains
unchanged and no obsolete eager-evaluation path is restored.

Validation

The exact nineteen-commit layer passes focused Go tests, integration compile,
locked Vitest, mandatory Linux lint/build, and Darwin MLX build. A fresh pinned
8B NVFP4 import passes the unchanged five-case release-media gate, including
image, audio, ordered history, MP3, metadata, semantic transcription, corrected
PNG count, unload, and empty /api/ps.

Full-stack evidence

The completed ten-layer v0.32.11 integration and release add-on passed their
owner-specific Linux, Darwin, import, inference, tool, thinking, media, and
unload gates. The signed/notarized preview DMG reports the exact packaged
version and passes its fresh six-source packaged-runtime matrix 6/6.

Limitations and out of scope

Apertus 1.1 Mini and affine checkpoints belong to #18.

AI assistance

Codex materially assisted with architecture adaptation, lifecycle validation,
tests, and evidence review. The layer received independent review before
acceptance.