feat(decoder): add support for granite models

Using straight Llama modeling for now -> generated content does not look
so good, so maybe differences in rope ?

Author: dacorvo

optimum/exporters/neuron/model_configs/decoder_configs.py

@@ -63,3 +63,9 @@ class Qwen2NeuronConfig(TextNeuronDecoderConfig):
     NEURONX_CLASS = Qwen2ForSampling
     CONTINUOUS_BATCHING = True
     FUSE_QKV = False
+
+
+@register_in_tasks_manager("granite", "text-generation")
+class GraniteNeuronConfig(TextNeuronDecoderConfig):
+    NEURONX_CLASS = LlamaForSampling
+    CONTINUOUS_BATCHING = True