Make swish compatible with nn.Sequential

[oiao]

nn.Sequential is used in multiple graph convolutional layers throughout the code. However, the current implementation of the swish activation is incompatible with this pattern:

from torch import nn
from torch_geometric.nn.resolver import activation_resolver
nn.Sequential(activation_resolver('swish'))

TypeError: torch_geometric.nn.resolver.swish is not a Module subclass

This PR replaces swish with class Swish(nn.Module).