Using LoRA, we allow to target nn.Parameters directly by using
target_parametes=[...]. This works by replacing the module that this
parameter belongs to with lora.ParamWrapper. However, this cannot work
if the parameter lives on the root model itself, since that would
require replacing the whole model. What actually happened was that the
module would self-reference, leading to an infinite recursion. With
this fix, we now detect this degenerate case and raise a proper error.

Note that I discovered this only by accident while working on a test
with a dummy model. It should be extremely unlikely that any real
world use cases would require targeting a parameter on the root
model. If we get such reports, we can think about a better solution.