[functional] torch.unique not implemented for booleans

[frgfm]

🐛 Bug

The torch.unique function is not implemented for torch.bool dtype.

To Reproduce

Steps to reproduce the behavior:

import torch
u1 = torch.unique(torch.rand(3,3), sorted=False)
u2 = torch.unique((10 * torch.rand(3,3)).round().to(dtype=torch.int), sorted=False)
u3 = torch.unique(torch.rand(3,3).round().to(dtype=torch.bool), sorted=False)

which produces

RuntimeError: "unique" not implemented for 'Bool'

Expected behavior

I understand the switch from torch.uint8 to torch.bool is fairly recent and that torch.unique does not seem very useful on booleans, but the feature is very useful when specifying the dim argument (torch.unique on boolean sequences for instance).

Environment

• PyTorch Version: 1.2.0
• OS: Ubuntu 18.04.3 LTS
• How you installed PyTorch: conda
• Python version: 3.7
• CUDA/cuDNN version: CUDA 10.1.168 (cuDNN 7.6.2)
• GPU models and configuration: GeForce GTX 1050 (driver: 430.50)

cc @ezyang @gchanan @zou3519 @jerryzh168 @izdeby

[cpuhrsch]

I'm flagging this as high priority because it is a regression in feature coverage when migrating from uint8 to bool tensors.

[ezyang]

cc @izdeby

[gchanan]

removing "triaged" because it's marked "high priority" but we didn't actually triage it.