Disable deterministic option in compile tests (#7720)
Compiler tests (with/without DeepCompile) occasionally fail with
mismatching loss values:
```
FAILED tests/unit/v1/compile/test_compile_zero.py::TestZeRO::test_compile_zero[none-1-dtype0]
AssertionError: Loss values are not close. Tensors are not close: actual=tensor(-0., device='cuda:1', dtype=torch.bfloat16, grad_fn=<DivBackward1>), expected=tensor(0.0255, device='cuda:1', dtype=torch.bfloat16,
grad_fn=<CompiledFunctionBackward>) kwargs={'rtol': 0.5, 'atol': 0.01}
```
While the exact root cause is not yet clear, but we found a [similar
issue](https://github.com/pytorch/pytorch/issues/159855) related to the
compiler.
This PR disables the deterministic option, which has improved stability.
Previously, we encountered this error intermittently when running the
compiler tests repeatedly. With this change, the tests now pass 100
consecutive runs.
---------
Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>
