feat: Improved attestation verification

[jcabrero]

I know this PR touches way too many files, but it wasn’t possible to break it into smaller PRs while keeping everything functional.

This PR introduces support for attestation verification using the official nvtrust library instead of the default Azure library.

To avoid issues caused by the outdated dependencies of nv-trust-attestation-sdk, we have modularized the attestation logic into a separate, standalone package that runs inside a Docker container. nilai-api now interacts with this container via an internal API, ensuring a cleaner and more maintainable integration.

To make it easier to review, here’s what you need to know:
1 .We have removed packages/verifier. This was a copy of the source from the Azure Confidential GPU library. Since the source was removed, that’s why there are so many lines of code deleted.
2. There is a new directory, nilai-attestation, which includes all the code required for attestation. It is a FastAPI server with a single endpoint at /attestation/report. It is intentionally kept completely separate from the rest of the nilai code to avoid any dependency conflicts. It is meant to run in the attestation: container in the Docker Compose setup and has its own attestation.Dockerfile.
3. Now nilai-api simply forwards attestations to the attestation container. It no longer needs to handle any verification itself, so it no longer has any GPU-related dependencies. Thus all the changes to nilai_api/state.py (which was in charge of the attestations before), and now there is nilai_api/attestation/__init__.py which forwards a GET request.
4. There are other QoL improvements, such as removing the two port bindings from the nilai-api gunicorn.conf.

[lumasepa]

This seems a bit hacky. I understand you are trying to load it only once. Have you tried to initialize it globally on import? It should be called only once. And you are not checking if the init failed swallowing the error and delaying it to the execution of the attestation.

Suggested change

	@lru_cache(maxsize=1)
	def load_sev_library() -> bool:
	"""Load the SEV library"""
	return sev.init()
	if not sev.init():
	raise Exception("Failed to load SEV")

[jcabrero]

At the moment, I aim at allowing nilAI to run on SEV and non-SEV hosts. If /dev/sev-guest is not available, the idea is that it continues to work, though with fake attestations. That's why we need to make sure an exception is not raised here.