test(model): drop MTP from the EXAONE MoE toy checkpoint

ExaoneMoeForCausalLM declares _keys_to_ignore_on_load_unexpected = [r"mtp.*"], so
released checkpoints carry MTP weights but the modelling class does not build them. A
toy constructed from that class with num_nextn_predict_layers set therefore saves no
MTP weights, while the provider still builds MTP layers and the bridge still registers
the MTP mappings off the same config value.

Those layers were being compared without ever having been loaded. The new mapping
validation reports them as missing instead, which is correct. Drop MTP from the toy so
it represents something the class can actually produce.

Verified with the roundtrip the functional test drives, TP=2: 89/89 weights matched.

Signed-off-by: Minho Ryu <ryumin93@gmail.com>

Author: bzantium

tests/functional_tests/test_groups/models/exaone/test_exaone_conversion.py

@@ -100,7 +100,11 @@
     "norm_topk_prob": True,
     "routed_scaling_factor": 2.5,
     "scoring_func": "sigmoid",
-    "num_nextn_predict_layers": 1,
+    # No MTP here. `ExaoneMoeForCausalLM` declares `_keys_to_ignore_on_load_unexpected =
+    # [r"mtp.*"]`, so released checkpoints carry MTP weights but the modelling class does
+    # not build them. A toy constructed from the class therefore saves nothing for the MTP
+    # mappings the bridge registers off `num_nextn_predict_layers`, while the provider still
+    # builds MTP layers, and the roundtrip would compare layers that were never loaded.
 }