TransformersModel.load_adapter: attach a PEFT adapter after construction

[JadenFiotto-Kaufman]

The post-hoc form of the peft= constructor kwarg, for #555.

model = TransformersModel("gpt2", task="text-generation")   # meta, no weights
model.load_adapter(adapter_id)   # grafts the adapter's config onto the meta tree
with model.trace(prompt):        # dispatch loads base + adapter weights together
    ...

Before dispatch, only the adapter's config is read: the meta module gains the adapter's architecture, so module paths (and remote requests) match the adapted model, and the real weights arrive with the base's at dispatch. This is what made the fallthrough PreTrainedModel.load_adapter fail there — safetensors cannot load onto meta.

After dispatch, it goes through the same swap the remote env path uses: call again with another id to swap adapters, or None to remove the current one, with the envoy tree rebuilt around the new structure each time.

Mechanically: the swap logic moves out of _remoteable_set_env into _swap_adapter, its rebind closure becomes _rebind, and the env path delegates to them unchanged. Our load_adapter deliberately shadows transformers' method of the same name, which mutates the module structure without the envoy tree noticing. Adapters attached this way ship with remote requests automatically, since _remoteable_get_env reads self.peft.

Four new tests in TestPeft against the local LoRA fixture; full suite green.

Closes #555

🤖 Generated with Claude Code