fix: add llm.base.pt fallback for V3 model switching

The LLM model switch logic only knew about llm.orig.pt as
non-RL fallback. V3 models (Fun-CosyVoice3-0.5B-2512) ship
with llm.base.pt instead of llm.orig.pt.

When use_rl_model=False, the code silently skipped the switch
because llm.orig.pt didn't exist, keeping llm.rl.pt active.
llm.rl.pt is incompatible with seed fallback audio, causing
the max_trials sampling crash.

Now fallback order: llm.orig.pt → llm.base.pt → no-op

aiia_cosyvoice_nodes.py

@@ -218,10 +218,14 @@ class AIIA_CosyVoice_ModelLoader:
         print(f"[AIIA] Loader Debug: use_rl_model_raw={use_rl_model} (type: {type(use_rl_model)}) -> effective_rl={is_rl_requested}")
         print(f"[AIIA] Loader Debug: llm_rl exists: {os.path.exists(llm_rl)}")
         
-        target_source = llm_rl if is_rl_requested and os.path.exists(llm_rl) else llm_orig
+        llm_base = os.path.join(model_dir, "llm.base.pt")
+        
+        # Determine the non-RL fallback: prefer llm.orig.pt, then llm.base.pt
+        non_rl_target = llm_orig if os.path.exists(llm_orig) else (llm_base if os.path.exists(llm_base) else None)
+        target_source = llm_rl if is_rl_requested and os.path.exists(llm_rl) else non_rl_target
         
         # 3. Aggressively Manage Symlink
-        if os.path.exists(target_source):
+        if target_source and os.path.exists(target_source):
             target_basename = os.path.basename(target_source)
             current_target = None