Fix module initialization with other dtypes and simplify module registration

[sbrunk]

Fixes and improves initializing modules with other parameter types.

import torch.*

// Default:
nn.Linear(10, 10) // Linear[Float32]

// Override explicitly:
nn.Linear[BFloat16](10, 10) // Linear[BFloat16]

// Override default via context parameter:
import Default.float64
nn.Linear(10, 10) // Linear[Float64]

This is currently inconsistent with tensor creation ops like torch.ones where we use a default parameter for the dtype, meaning it's easier to override at runtime, but we have a fixed default. Both designs have tradeoffs and we need to test what works well in practice. Perhaps we can event find a way to combine both approaches.

Type of all modules fixed at creation time

To make container classes like Sequential work, we currently fix the dtype on module initialization even for parameterless modules like Softmax:

final class Softmax[D <: DType: Default](dim: Int) extends TensorModule[D]:
  def apply(t: Tensor[D]): Tensor[D] = Tensor(nativeModule.forward(t.native))

Instead of a generic apply method:

final class Softmax(dim: Int):
  def apply[D <: DType](t: Tensor[D]): Tensor[D] = Tensor(nativeModule.forward(t.native))

If we're able to find a better way, we might be able to change this back in the future.

DType conversion of modules

Module type conversion will need rethinking the module design taking into account (im)mutability of modules and things like mixed precision training and perhaps quantization. For now the type is fixed on module creation.

Other changes

• Simplify module registration after improvements in new presets.
• Remove broken copy and to(dtype) methods from module.