GPU TEE attestation: verify GPU confidential mode and bind it to the CVM (currently unverified)

## Problem

dstack attaches NVIDIA GPUs to CVMs but never verifies them. The CPU-side stack has no GPU/NRAS logic, and the guest OS flips the GPU to "ready" **unconditionally** at boot, before the workload runs. Consequences:

- A **non-confidential GPU** (CC off in host BIOS, or a normal GPU) is exposed **plaintext** to the workload with no error that stops it — no attacker required.
- Even a genuine CC GPU is **not bound to the CVM**: nothing lets a relying party (or the KMS) trust that a specific attested CVM is backed by a genuine GPU in confidential mode.

"Dual attestation" in `docs/security/security-model.md` is a documented *expectation*, not something enforced anywhere.

Full design write-up: **`docs/gpu-attestation-design.md`** (branch `gpu-tee-nras-verification`).

## Current gaps (evidence)

- No CPU-side GPU verification — `AttestationQuote` is CPU-TEE only; `VerificationDetails` has no GPU fields (`verifier/src/verification.rs`).
- GPU attached as plain `vfio-pci`, no CC-mode/attestation (`vmm/src/app/qemu.rs:840`).
- `nvidia-smi conf-compute -srs 1` run unconditionally at boot, no `cc_mode` check (meta-dstack `nvidia-persistenced.service`).
- `app-compose.service` has **no dependency** on GPU bring-up — the workload starts even if it fails.
- No app-facing RPC to query/enable GPU confidentiality.

## Threat model (what this closes / doesn't)

| Attack | Closed? |
|---|---|
| Forge GPU evidence (no NVIDIA key) | ✅ cert chain + RIM, verifier in measured guest |
| Non-CC / CC-off GPU silently used | ✅ fail-fast, `cc_mode==ON` |
| **Copy attack** — GPU-less instance B copies a GPU-attested value from A | ✅ only if verdict lives in append-only measured pre-app state; ❌ if in `report_data` (app-forgeable via `guest-agent/src/rpc_service.rs:327`) |
| Stale/replayed GPU evidence | ✅ fresh boot nonce |
| **Live relay / cuckoo** to a genuine remote CC GPU | ❌ residual — no shipping system defeats it; needs TEE-IO/TDISP hardware. Document, don't imply "GPU proven local." |

## Workstreams

### 1. Offline local verifier (NVAT)
- [ ] Adopt NVIDIA C++ Attestation SDK (`libnvat`, [NVIDIA/attestation-sdk](https://github.com/NVIDIA/attestation-sdk)); Python `nvtrust` is EOL 2026-09-15.
- [ ] Bake `libnvat` + pre-provisioned filesystem RIM store into the NVIDIA image (re-provision on driver/VBIOS upgrade).
- [ ] Handle OCSP (the one online dep, `ocsp.ndis.nvidia.com`): in-CVM caching/replay proxy at `--ocsp-url`, and/or a Rego policy tolerating `x-nvidia-cert-ocsp-status`.

### 2. Binding via measured append-only state (before the app boundary)
- [ ] `dstack-util` emits a `gpu-attestation` event committing `H(nvat_eat‖cert_chain‖claims)` **before** `system-ready`.
- [ ] Verifier trusts it via the existing `find_event` boundary (breaks at `system-ready`, `dstack-attest/src/attestation.rs:176`) → RTMR3-bound on TDX.
- [ ] `report_data` stays for freshness + RA-TLS key binding only — **not** the GPU verdict.

### 3. Fail-fast enforcement & app gate
- [ ] `dstack-gpu-attest.service`: verify every configured GPU (`num_gpus>0` is measured), require `cc_mode==ON` (reject `OFF` **and** `DEVTOOLS`), set `-srs 1` only on pass.
- [ ] `app-compose.service` `Requires=`+`After=dstack-gpu-attest.service`; fail closed (no `system-ready`, no workload).
- [ ] Guest-agent RPC `GetGpuAttestation()` / `EnsureGpuReady()` so apps can confirm/enable before use.

### 4. KMS gating, verifier surfacing, docs
- [ ] KMS gates key release on the `gpu-attestation` event (like `compose-hash`).
- [ ] Verifier parses NVAT EAT/claims into `VerificationDetails`.
- [ ] `security-model.md`: state the guarantee (*measured-guest-vouches + channel-bound*) and the co-location residual.

## Caveats

- **SEV-SNP is blocked**: no runtime measurement register (`tpm_runtime_pcr()=None`, `has_tdx()=false`); `decode_app_info_sev_snp` reads identity from launch-time `HOST_DATA`/`MrConfigV3` and ignores the runtime log. SNP needs an **SVSM/coconut-vTPM** (PCR channel) before this works — until then SNP GPU attestation is strictly weaker than TDX. (See #713.)
- **DEVTOOLS** CC mode enables CC APIs *without* memory encryption — must be rejected.
- **Co-location residual** stands until Blackwell TEE-IO/TDISP.

## Hardware validation checklist (unconfirmed in docs)

- [ ] NVAT behavior on OCSP connection failure (fail-open vs closed); can Rego neutralize revocation?
- [ ] OCSP `nextUpdate` window; warm-cache ride-through.
- [ ] Exact `libnvat` env-var spellings; trust-root store overridability.
- [ ] Does `-srs 1` error or no-op on a CC-off GPU?
- [ ] Does the driver expose SPDM `KEY_EXCHANGE` key (optional session-binding stretch)?
- [ ] SNP SVSM vTPM feasibility in meta-dstack.

Start with docs/gpu-attestation-design.md, then inspect verifier/src/verification.rs, vmm/src/app/qemu.rs:840, dstack-attest/src/attestation.rs:176, and guest-agent/src/rpc_service.rs:327. Trace the existing measured-event boundary and GPU boot services before choosing a workstream. Done requires fail-closed GPU verification and binding, with verifier, KMS, RPC, service-gating, and documentation changes covered as applicable.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.