Enable bfloat16 for hardtanh_backward_cuda (#91511)
I'm not sure why this was left out in the first place as all adjacent operations have both Half and BFloat16. Things seem to work as expected and this enables `relu6` to be used in bfloat16 training. Hardtanh backward is super simple and precision is not relevant.
```
import torch
x_fp32 = torch.tensor([-1,2,4,7], requires_grad=True, dtype=torch.float32, device="cuda")
x_bf16 = torch.tensor([-1,2,4,7], requires_grad=True, dtype=torch.bfloat16, device="cuda")
torch.nn.functional.relu6(x_fp32).sum().backward()
torch.nn.functional.relu6(x_bf16).sum().backward()
assert (x_fp32.grad == x_bf16.grad).all()
```
Previously would fail with:
```
Traceback (most recent call last):
File "test_hardtanh_patch.py", line 5, in <module>
torch.nn.functional.relu6(x_bf16).sum().backward()
File ".../lib/python3.8/site-packages/torch/_tensor.py", line 396, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File ".../lib/python3.8/site-packages/torch/autograd/__init__.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: "hardtanh_backward_cuda" not implemented for 'BFloat16'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91511
Approved by: https://github.com/ngimel
