`chalf` is experimental as the op coverage is low.

Following script raises 6 warnings if `set_warn_always(True)` else raises only 1 warning.
```python
import torch
torch.set_warn_always(True)
device='cpu'
t = torch.randn(3, dtype=torch.chalf, device=device)
y = torch.rand(3, dtype=torch.chalf, device=device)
# Allocates new tensor for result
t + y

device='cuda'
t = torch.randn(3, dtype=torch.chalf, device=device)
y = torch.rand(3, dtype=torch.chalf, device=device)

# Allocates new tensor for result
t + y

```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78245
Approved by: https://github.com/anjali411