feat: Support Flux2 in other common components

[wang-shuibin]

This PR implements adaptation for Flux2 in common components. Meanwhile, accuracy verification for the Flux2 model has been finished Flux2 model.

[gemini-code-assist]

Code Review

This pull request implements a Mistral prompt-to-message conversion flow and extends the Tokenizer interface to support a max_sequence_length parameter with padding logic. Feedback highlights several critical issues: a potential out-of-bounds crash and debug code in EmbedWorkerImpl, a violation of the repository style guide regarding global flag registration in help_formatter.h, and multiple safety concerns in LLMMaster including missing error callbacks, exception-unsafe JSON parsing, and the failure to clear stale prompt tokens after a prompt update.

[wang-shuibin]

This PR adapts common components for Flux2 and completes accuracy verification for the Flux2 model. Additionally, it fixes a bug in xllm/core/layers/common/add_matmul.cpp.