🐛 Describe the bug

Output of arange bf16 is not correct. use below code to reproduce

torch.arange(241,273, dtype=torch.bfloat16)

Output:

tensor([241., 242., 243., 244., 245., 246., 247., 248., 249., 250., 251., 252., 253., 254., 255., **256., 256., 256**., 258., 260., 260., 260., 262., 264., 264., 264., 266., 268., 268., 268., 270., 272.], dtype=torch.bfloat16)

Here after 255 the results are shown as [256, 256, 256] which is wrong as in f32 output is [256, 257, 258] but 257 is not representable in bf16 hence it should be [256, 256, 258]

Versions

torch==2.4.1

### Tasks

cc @svekars @brycebortree @sekyondaMeta @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @albanD