phase_advance should be a buffer so it moves device correctly

[ksanjeevan]

@vincentqb, fixing what was mentioned #456.

[vincentqb]

Can you show an error that would occur without that change? Some attributes/etc are inferred by nn.Module, see here.

[ksanjeevan]

Yeah so here's an example:

import torch
from torchaudio.transforms import TimeStretch
m = TimeStretch().cuda()
m(torch.randn(1, 2, 201, 400, 2).cuda(), 1.3)

Will give a RuntimeError: expected device cuda:0 but got device cpu. Also we're using buffers on the other transforms as far as I'm aware, see here for example.

[vincentqb]

Yup, I can reproduce the error. Good catch!