GKernelSE2 returns nan kernel when using rbf as sampling mode

[dgcnz]

Reproduction

For parameters: grid_H=None, group_kernel_size=4, this will return a group_sampling_width=0 which will in turn make rbf return a tensor of nan's.

gconv/gconv/gnn/kernels/kernel_sen/kernel_se2.py

Lines 233 to 239 in 4879da6

	if grid_H is None:
	grid_H = so2.uniform_grid(group_kernel_size)
	grid_Rn = gF.create_grid_R2(kernel_size)
	if not group_sampling_width:
	group_sampling_width = 0.8 * so2.nearest_neighbour_distance(grid_H).mean()

Temporary fix

Use group_sampling_mode='nearest'.