Update sample inputs for fft.hfftn (#81416)
Our internal runs show that `test_complex_half_reference_testing_fft_hfftn_cuda_complex32` is flaky, and fails quite often.
One way to replicate the test:
```
import torch
def allclose(a, b, rtol, atol):
return torch.all(torch.abs(a-b) < atol + rtol * torch.abs(b)).item()
failed = 0
N = 100
for i in range(N):
inp = torch.rand((2, 2, 65), dtype=torch.complex32, device ='cuda')
out_complex32 = torch.fft.hfftn(inp)
out_complex64 = torch.fft.hfftn(inp.to(torch.complex64))
failed = failed + int(allclose(out_complex32, out_complex64, rtol=0.04, atol=0.04))
print("test failed {} times out of {} times".format(failed, N))
```
I suggest to decrease the size of the input tensor in `sample_inputs_spectral_ops` for `hfftn` to avoid failing.
cc @kshitij12345
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81416
Approved by: https://github.com/peterbell10, https://github.com/ngimel
