Replace torch.linspace with np.linspace

[fepegar]

This allows to initialize models in the "meta" device, which is >10× faster, if desired.

import torch

from dynamic_network_architectures.architectures.primus import Primus

def initialize_primus():
    return Primus(1, 864, 3 * (8,), 2, input_shape=3 * (160,))

%%timeit
initialize_primus()

2.45 s ± 38 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%%timeit
with torch.device("meta"):
    initialize_primus()

195 ms ± 871 μs per loop (mean ± std. dev. of 7 runs, 1 loop each)

[fepegar]

FYI @TaWald @FabianIsensee

[fepegar]

Is this change welcome? Otherwise I'll start working from a fork.

[TaWald]

Fabi is currently away for the next two weeks and AFAIK he is the only person who has permissions to merge these changes into the official MIC-DKFZ repo. Once he's back I can ping him to merge this

[fepegar]

Thanks, @TaWald.

[fepegar]

Is it possible to review, merge and release? I'm happy to help with maintenance if needed.