Eliminate import side effects

[jfeser]

The following program should work without needing a dummy import to run the side effects for effectful.handlers.torch.

import torch
from effectful.ops.syntax import defop

x = defop(torch.Tensor)

t = torch.randn((2, 3))[x()]

[eb8680]

I think this is true for number types (resolved in #335), but the reason we defer the imports of handlers.torch and handlers.jax is that those two libraries themselves have import side effects that don't always play nicely with each other. In particular, XLA likes to allocate all GPU memory for itself immediately, even if you aren't actually doing a JAX computation.

[jfeser]

I was hopeful that some kind of lazy import (https://peps.python.org/pep-0690/) could help, but unfortunately singledispatch functions require the whole module to be imported so they can access the class objects. I don't see a path towards fixing this.