Build UNSLOTH_MODEL_NAME fresh per load (fixes gpt-oss merged reload)

UNSLOTH_MODEL_NAME was built by prepending its previous os.environ value, so it
accumulated load flags across loads and across processes (it is inherited by a
save->reload subprocess). After loading a gpt-oss bnb-4bit model, a leftover
_load_in_4bit_ stayed in the string, and a subsequent non-4bit load (e.g. reloading a
merged_16bit checkpoint) still saw _load_in_4bit_. gpt-oss then kept its BnB router
patch (router.linear.weight) while the merged checkpoint stores the stock router.weight,
raising 'Unsloth: Critical error since some weights are not initialized'.

Rebuild the string from THIS load's model name and flags so each load is self-contained.
The model-type tag and load flags are still present, so all UNSLOTH_MODEL_NAME consumers
keep working; this only drops stale flags carried over from an unrelated earlier load.

unsloth/models/loader.py

@@ -1234,7 +1234,14 @@ class FastModel(FastBaseModel):
 
         # Save model types and loading method
         lowered_model_name = model_name.lower()
-        string = os.environ.get("UNSLOTH_MODEL_NAME", "") + model_types_all
+        # Build UNSLOTH_MODEL_NAME freshly from THIS load's model name + flags. Do not
+        # prepend the previous os.environ value: it persists in-process and is inherited
+        # across processes (e.g. a save->reload subprocess), so it accumulates stale load
+        # flags. A leftover "_load_in_4bit_" from an earlier bnb-4bit load would make
+        # gpt-oss wrongly take the BnB router patch (router.linear.weight) when later
+        # reloading a merged 16bit checkpoint (router.weight), raising
+        # "Unsloth: Critical error since some weights are not initialized".
+        string = lowered_model_name + "," + model_types_all
         if load_in_4bit:
             string += "_load_in_4bit_"
         if load_in_8bit: