simplify `AutoSoftConditioningReparam` to use `soft_eq` directly.

[rfl-urbaniak]

This is a follow-up to PR #500 .

The strategy:

1. Given a constraint and a scale/alpha passed as arguments to AutoSoftConditioning define _soft_eq that uses these values, and pass it on to KernelSoftConditionReparam, with proper constraint diagnosis preceding this call.

2. Eliminate TorchKernel, SoftEqKernel and RBFKernel completely.

3. Simplify AutoSoftConditioning accordingly.

Remark. I'm not sure if:

scale = self.scale * functools.reduce(
                 operator.mul, msg["fn"].event_shape, 1.0
             )

is still needed, commented out for now.

4. Modify tests/observational/test_handlers.py accordingly, to avoid using these kernels.

[rfl-urbaniak]

This is sufficient for tests/observational/test_handlers.py to pass, but tests/dynamical/test_handler_composition.py fails.

1. Reinstated

scale = scale * functools.reduce(
               operator.mul, msg["fn"].event_shape, 1.0
            )

still, same failure.

2. Considered reverting to the intermediate variant that preserves TorchKernel as a torch.nn.Module, same failure.

@SamWitty let's chat about this at some point.