[Docs] Add code generation architecture documentation

[tlopex]

This pr adds an architecture document explaining how PrimFuncs are compiled to executable code. Covers the codegen's position in the tvm.compile() pipeline, the target dispatch mechanism (target.build.<kind> FFI lookup), the two codegen families

[gemini-code-assist]

Code Review

This pull request introduces a new architectural document, docs/arch/codegen.rst, which provides a comprehensive overview of the TVM code generation pipeline, including target dispatch mechanisms, the LLVM and Source codegen families, and the runtime module system. The documentation is also integrated into the main architecture index. The review feedback identifies several areas for improvement: correcting the C++ code snippet to use standard TVM runtime APIs instead of internal FFI types, fixing an incorrect file path for the CUDA build flow, and adding a missing entry for WebGPU in the source code map to ensure completeness.