Better convolve1d implementation on jax and torch

[hexane360]

• On jax, convolve1d uses conv_general_dilated directly, which should dispatch to a 1D convolution rather than a 3D convolution with kernel size (1, 1, N).
• On torch, xp.pad has a fallback for cases torch.nn.functional.pad doesn't support
• On torch, convolve1d now uses xp.pad. Importantly, kernel size >= object size now works
• Add test cases for xp.pad

[hexane360]

@mlz-EM can you review when you get a chance? Would be helpful if you could benchmark layers regularization on jax before and after, make sure it's not significantly slower