Add optimizing-models-with-quark entry-point skill

[amd-zhaofeng]

What

Adds optimizing-models-with-quark, an entry-point skill that connects an agent to AMD Quark's own quantization skills. Implements #139.

Why

amd-skill-finder already routes "quantize this model to FP8" to amd/Quark, but all it can return is a repository address — the agent still has to read the docs and guess the flags. Quark ships skills of its own for exactly this work, and catalog users cannot reach them.

The skill is deliberately thin: it quantizes nothing, installs no pip package, and registers nothing as a skill. The quantization knowledge stays upstream in Quark, versioned with the product, so there is no vendored copy here to go stale.

How it works

1. Check whether Quark's skill tree is already in the cache directory (~/.cache/amd-quark-skills, or %LOCALAPPDATA%/amd-quark-skills on Windows).
2. If not, explain what will happen and fetch it with the user's consent — a shallow, blobless, sparse clone, nothing written to their workspace.
3. Verify the starting-point files are readable; stop and report if not.
4. Route to one of three starting points and follow it, checkpoints included.

The user wants to	Starting point
Quantize a PyTorch / HuggingFace / safetensors model	quark-torch-llm-ptq-workflow
Quantize an .onnx model	quark-onnx-ptq-workflow
Install or verify the Quark environment	quark-install

Interim design for amd-quark v0.12

The proposal in #139 assumed pip install amd-quark would deliver the skills.
On amd-quark v0.12 it does not, so this PR takes a different route and says so plainly:

• The pip package does not carry the skills. setup.py packages quark and quark.* only; the skills live in the repository, so they are fetched from GitHub instead.
• A skill is not self-contained. .claude/skills/ holds thin stubs that defer to .claude/skills-impl/, whose files reference each other by relative path. The tree therefore travels as a unit — fetching individual files is not reliable.
• Read as files, not installed as skills. Nothing is written into the user's workspace or any harness skills directory, and no restart is required. The trade-off is stated in the skill body: because nothing is registered, a later question will not route to a Quark skill on its own and every request goes through this entry point. That cost is temporary, and it buys a clean removal path — a single rm -rf of the cache directory.

Once Quark ships its skills inside the quark package, the retrieval step goes away and this entry point can become what #139 originally described.

Test plan

• ./.github/scripts/check.sh passes — 8 skills, 0 errors, both generated manifests report up to date
• Fetch commands smoke-tested end to end: git clone --sparse plus sparse-checkout set yields all four required files, ~2 MB on disk
• Description is third person and 824 characters, with negative triggers for serving-llms-on-instinct / serving-llms-on-epyc and magpie-kernel-evaluator
• SKILL.md body is 164 lines, well under the 500-line cap; reference is linked one level deep
• Not included this round: evals/evals.py, a walkthrough, and routing eval prompts — happy to add in a follow-up