Add Ernie 4.5 MOE and 0.3B Support

[firecoperana]

This is the port of Ernie 4.5 MoE model and Ernie 4.5 0.3B
In my test, response from the model looks ok.

Didn't test the python convert script. It's a straight copy/paste.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[Ph0rk0z]

This works for big ernie too?

[firecoperana]

Yes.

[AesSedai]

The python convert script is broken, because of the following error:

Traceback (most recent call last):
  File "/home/jarvis/development/ik_llama.cpp/convert_hf_to_gguf.py", line 2198, in <module>
    @ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")
NameError: name 'ModelBase' is not defined

The mainline may use ModelBase, but in ik_llama it's just @Model.register

Edit: also, TextModel should just be Model in class Ernie4_5Model(Model):

[firecoperana]

Can you verify if #774 fix the error?

[AesSedai]

@firecoperana confirmed that PR un-breaks the convert script. I haven't tried an ERNIE convert, I was doing a GLM-4.5 quant and just noticed that the script was totally busted, that PR fixes the issue I was seeing.