`torch_xla::runtime::PjRtComputationClient::ExecuteReplicated()` crash after RuntimeError

See crash at: https://github.com/tengyifei/playground/blob/master/torch-xla-device.ipynb

Minimal repro:

```python
import torch
import torch_xla

a = torch.tensor([1.0], device="xla:1")
b = torch.tensor([1.0], device="xla:2") 
print(a.device, b.device)

a + b

a + b.cpu()

import torch_xla.runtime as xr
xr.use_spmd()

c = torch.tensor([1.0], device="xla") 
c.device, c + c
```


