Configure tests on RewardFunction monotonicity, vmap consistency to test performance both on CPU and GPU (where we have some non-determinism)

[k-chrispens]

[Abdelsalam-Abbas]

The cleanest approach here is to parametrize the device fixture to yield both cpu and cuda, then set tolerances conditionally based on device type:

atol = 1e-6 if device.type == "cpu" else 5e-4
torch.testing.assert_close(result_vmap, result_sequential, rtol=0, atol=atol)

CPU runs catch real logic bugs with tight tolerances, GPU runs validate it works under floating-point non-determinism. This avoids duplicating tests.

Tolerance fix for the immediate GPU failures is in #173.