clarify the "adapting when a LoRA is injected" issue in tests

tests/adapters/test_lora.py

@@ -45,6 +45,15 @@ def test_lora_adapter() -> None:
     a2 = LoraAdapter[fl.Chain](chain, sub_targets=chain.walk(fl.Linear), rank=1, scale=1.0).inject()
     assert len(list(chain.layers(Lora))) == 6
 
+    # If we init a LoRA when another LoRA is already injected, the Linear
+    # layers of the first LoRA will be adapted too, which is typically not
+    # what we want.
+    # This issue can be avoided either by making the predicate for
+    # `walk` raise StopIteration when it encounters a LoRA (see the SD LoRA)
+    # or by creating all the LoRA Adapters first, before injecting them
+    # (see below).
+    assert len(list(chain.layers(Lora, recurse=True))) == 12
+
     # ejection in forward order
 
     a1.eject()
@@ -60,6 +69,10 @@ def test_lora_adapter() -> None:
     a2.inject()
     assert len(list(chain.layers(Lora))) == 6
 
+    # If we inject after init we do not have the target selection problem,
+    # the LoRA layers are not adapted.
+    assert len(list(chain.layers(Lora, recurse=True))) == 6
+
     # ejection in reverse order
 
     a2.eject()