Abate spurious resize warnings in `MultiMarginLoss` on CUDA
The current `multi_margin_loss_cuda_out` implementation has a few cases where an output is intended to have shape `{0}` before resizing but is improperly set before this step, triggering `UserWarning`s. This PR changes the initial `at::empty` call to avoid this warning and alters the current resizing logic to match that of the CPU version to avoid the warning when calling `at::sum_out`.
The original issue was reported from the user forum by @hadaev8 [here](https://discuss.pytorch.org/t/warning-spam-then-using-multimarginloss/147492).
CC @ptrblck
Original repro provided by @hadaev8:
```
import torch
from torch import nn
from torch.nn import functional as F
bs = 56
model = nn.Linear(128, 22).cuda()
loss = nn.MultiMarginLoss()
x = torch.rand((bs, 128)).cuda()
targets = torch.randint(22, (bs,)).cuda()
out = model(x)
print(targets.shape)
print(out.shape)
loss(out, targets)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75000
Approved by: https://github.com/ngimel
