bugfix: init ssm_cache by config ssm_cache_type and fix dimension order of g & beta for qwen3.5.

[JC-ut0]

This PR makes the following changes to the xLLM inference engine:

1. Add mamba_ssm_dtype config field — A new mamba_ssm_dtype string property is added to ModelArgs , allowing the SSM (State Space Model) cache dtype to be specified independently from the model's primary dtype.
2. Use mamba_ssm_dtype for KV cache capacity estimation — In llm_engine.cpp , the estimate_kv_cache_capacity() function now uses the mamba_ssm_dtype -derived byte size for calculating the SSM slot size, instead of always using the model dtype size.
3. Fix g/beta tensor layout in GDN gating — In qwen3_gated_delta_net_base.cpp , after calling fused_gdn_gating() , the g and beta tensors are permuted from [seq_len, batch, heads] to [batch, seq_len, heads] layout via .permute({1, 0, 2}).contiguous() .
4. Load mamba_ssm_dtype from model config — The qwen3_5.h model registration macro now loads mamba_ssm_dtype from the JSON config (with text_config.mamba_ssm_dtype fallback) using LOAD_ARG_TEXT_OR_ROOT .

[gemini-code-assist]

Code Review

This pull request implements the LLMEngine and WorkerImpl components, providing the core infrastructure for distributed inference, KV cache management, and model lifecycle operations. It adds support for advanced architectures like DeepSeek v3 and Qwen3, along with optimizations such as XTensor memory management and rolling weight loading. The review feedback suggests enhancing type safety by using enums instead of string literals for model arguments. Additionally, the implementation should be refined to adhere to the project's style guide, particularly concerning fixed-width integer usage, constant naming conventions, and the appropriate use of the auto keyword.

[JC-ut0]

/gemini-review

[JC-ut0]

/gemini-review

[yingxudeng]

Great catch.

[yingxudeng]

CUDA 还没编译完，但是 非NPU的很多已经编译通过了。这个pr已经排队两天了，先合并吧，应该cuda应该编译没问题