feat(service): implement interactive chat loop in qwen3 service

[chenghuaWang]

• Replace hardcoded request with interactive user input loop
• Add support for multi-turn conversations with history tracking
• Integrate thinking state visualization using fmt library
• Handle graceful exit with /exit or /quit commands
• Improve response formatting with proper JSON structure

fix(cpu): support BSHD layout in RoPE operation

• Extend RoPE operator to handle both BHSD and BSHD tensor layouts
• Add layout type checking and dimension mapping logic
• Fix assertion to allow multiple input layout types

refactor(service): improve request pool shutdown handling

• Change RequestPool::pop() to return std::optional
• Update service worker loop to handle optional requests
• Ensure worker threads join properly during shutdown
• Add timestamp to response payloads
• Modify response format to follow chat completion chunk structure

perf(cache): optimize ZenFS blob size calculation

• Simplify bit calculation using literal 64 instead of sizeof(uint64_t)
• Correct blob size computation by using consistent dtype lanes

fix(qwen3): correct loop initialization and template processing

• Initialize loop counters properly in attention mechanisms
• Refine chat template processing for better tool and thinking support
• Adjust tensor shapes for sequence and position IDs
• Update default sampling parameters and EOS token ID
• Change cache data types from float16 to float32 for better precision

style(qwen3): reformat function signature for readability

• Break long function declaration into multiple lines
• Improve code formatting and parameter alignment