docs(skills): add 9 task-focused skills under skills/usecase/

[Marvin-Cypher]

Summary

Adds 9 task-focused skill markdowns under `skills/usecase/` that AI coding agents (Claude Code, Cursor, Codex) fetch to scaffold Phala Cloud projects end-to-end. Builds on the foundational `skills/phala-cli/SKILL.md` (CLI reference).

All commands validated against live `phala` CLI v1.1.9 + live API (`cloud-api.phala.com/api/v1/attestations/verify` returned `verified: true` with matching `mr_config` ↔ `compose_hash` against a real running CVM).

Skills

Skill	What it covers
`agent-deploy.md`	Confidential AI agent — sealed credentials via `-e`, RA-TLS, Sign-RPC action log, multi-agent fleet
`gpu-vllm-deploy.md`	Self-hosted vLLM on H200 GPU TEE — model catalog, multi-GPU tensor-parallel, GPU CC verification
`inference-call.md`	Call the hosted Confidential AI API at `api.redpill.ai/v1` — cURL/Python/TS, streaming, tool calling
`training-run.md`	SFT / DPO / LoRA / continued PT on GPU TEE with TRL or Unsloth — sealed dataset, signed manifest
`data-coanalysis.md`	Multi-party cohort analysis — owner sealing, multi-sig DstackApp, prepare-only/commit, DP-aggregate output
`gpu-tee-custom.md`	Generic custom workload on GPU TEE — Jupyter quickstart + common Docker images
`dstack-self-host.md`	Self-host `dstack-vmm` + `dstack-kms` + `dstack-gateway` on bare-metal TDX with `vmm-cli.py`
`cloud-migration.md`	Migrate from AWS Nitro / GCP CC VM / Tinfoil — diff map, shadow traffic, cutover
`verify-attestation.md`	End-to-end TEE verification — NVIDIA NRAS + Intel TDX + report-data binding + compose-hash + Sigstore

Folder layout

```
skills/
├── phala-cli/ ← existing CLI reference (SKILL.md)
│ ├── SKILL.md
│ └── references/
└── usecase/ ← new: task-focused recipes (9 files)
├── agent-deploy.md
├── gpu-vllm-deploy.md
├── inference-call.md
├── training-run.md
├── data-coanalysis.md
├── gpu-tee-custom.md
├── dstack-self-host.md
├── cloud-migration.md
└── verify-attestation.md
```

The use-case skills cross-reference `../phala-cli/SKILL.md` for foundational steps (install, login, debug, SSH).

Use case

Marketing pages on `phala.com` will embed 2-line "vibe-code" prompts pointing at the raw URLs:

```
Set up a confidential AI agent on Phala Cloud.
Fetch https://raw.githubusercontent.com/Phala-Network/phala-cloud/main/skills/usecase/agent-deploy.md
and follow the steps. Confirm with `phala cvms attestation `.
```

The user pastes this into their AI coding agent, which fetches the skill and walks through scaffolding + deploy + verification.

Format

Each skill matches the existing `SKILL.md` convention:

• YAML front-matter with `name` + multi-line `description`
• H1 title + 1-line tool intro
• `## Operations` decision matrix
• Task sections as `## TaskName` with numbered `### Step N` substeps
• Bash code blocks with comments
• Reference tables for instance types / models / failure modes
• Cross-references each other (e.g., `agent-deploy.md` references `verify-attestation.md` for the full verification flow)

Test plan

• Each `Reference: minimal end-to-end` snippet at the bottom of each skill copy-pastes cleanly into a terminal
• Cross-references between skills resolve once merged to `main`
• `verify-attestation.md` minimal Python end-to-end runs against a live deploy

🤖 Generated with Claude Code