Add GPU tests to vector field methods

[StarostinV]

Extend tests/inference_on_device_test.py for checking inference and training on GPU for unified vector field methods after merging PR #1497 and #1501 .

[jorobledo]

PR #1527 includes GPU tests for sampling from vector field posteriors

sbi/tests/inference_on_device_test.py

Lines 703 to 726 in d693840

	@pytest.mark.gpu
	@pytest.mark.parametrize("device", ["cpu", "gpu"])
	@pytest.mark.parametrize("iid_method", ["fnpe", "gauss", "auto_gauss", "jac_gauss"])
	@pytest.mark.parametrize("inference_method", [FMPE, NPSE])
	def test_VectorFieldPosterior(
	device: str, iid_method: str, inference_method: VectorFieldInference
	):
	device = process_device(device)
	prior = BoxUniform(torch.zeros(3), torch.ones(3), device="cpu")
	inference = inference_method(score_estimator="mlp", prior=prior)
	density_estimator = inference.append_simulations(
	torch.randn((100, 3)), torch.randn((100, 2))
	).train()
	posterior = inference.build_posterior(density_estimator, prior)
	posterior.to(device)
	assert posterior.device == device, (
	f"VectorFieldPosterior is not in device {device}."
	)
	x_o = torch.ones(2).to(device)
	samples = posterior.sample((2,), x=x_o, iid_method=iid_method)
	assert samples.device.type == device.split(":")[0], (
	f"Samples are not on device {device}."
	)

Would this be enough or do you think something else needs testing?

[janfb]

We probably also want to train on the device, so pass

inference = inference_method(score_estimator="mlp", prior=prior, device=device)

[StarostinV]

@jorobledo thanks a lot for the gpu tests! I think it would be nice to also add another test to check log_prob on GPU (with iid_method=None since it iid not supported yet until the corresponding PR #1508 is merged) - note that it's quite slow with the default settings, which aren't needed since we don't test the quality of log_prob here, so one can accelerate it as it is done here linearGaussian_vector_field_test.py:

# faster but inaccurate log_prob computation
posterior.potential_fn.neural_ode.update_params(
    exact=False,
    atol=1e-4,  # maybe even 1e-3
    rtol=1e-4,  # same
)
log_prob = posterior.log_prob(thetas)

Also, I think there is no need to test NPSE and FMPE here separately since they share log_prob implementation