Minicpm enabling

[pi314ever]

What does this PR do?

Enables MiniCPM3 model for Causal LM. Follows #1133 in optimizing remote code. The following changes were added:

• htcore.mark_step() between each decoder layer to reduce graph size
• Add token_idx support passthrough to Attention implementations. (Currently incomplete, missing reset of cache)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[pi314ever]

Performance benchmark
Command:

python run_generation.py --model_name_or_path openbmb/MiniCPM3-4B  --max_new_tokens 512 --prompt "推荐5个北京的景点。" --batch_size 1 --temperature 0.7 --do_sample --top_p 0.7 --use_kv_cache --use_chat_template --use_hpu_graphs [--bf16] 

throughput (tokens/s)	A100	Gaudi 2 (Graph mode)
BF 16	20.09	65.16
FP 32	23.89	44.46

[libinta]

can you list what the change in each section?

[pi314ever]

The files with "Copied from..." did not have any changes, while the files with "Adapted from..." have changes. In modeling_minicpm.py, I added docstrings for each method that I changed with what I modified. Is this sufficient? Or should I add a generic summary on the file docstring as well?

[yao-matrix]

@pi314ever , you need to add jsonschema and datamodel_code_generator into requirements. Or remove custom tokenizer since the model owner already switch back to Llama tokenizer, as here https://huggingface.co/openbmb/MiniCPM3-4B/commit/e8a65f63cd4e4eff91571e603a2a34e50628ff67#d2h-846292

[pi314ever]

@yao-matrix I have removed the tokenizer and validated it to run with default requirements.txt. There are no performance differences from basic testing.