[Docs] Add BYOC external library dispatch architecture documentation

[tlopex]

Add an architecture document (docs/arch/external_library_dispatch.rst) for TVM's BYOC (Bring Your Own Codegen) mechanism, covering the end-to-end pipeline for offloading operator subgraphs to external libraries (cuBLAS, CUTLASS, cuDNN, DNNL)

[gemini-code-assist]

Code Review

This pull request introduces a new architectural document, external_library_dispatch.rst, which provides a detailed overview of the External Library Dispatch (BYOC) pipeline in TVM Relax. The documentation covers pattern registration, partitioning, code generation, and runtime execution, and is integrated into the main architecture index. Feedback focused on technical accuracy: correcting the pipeline diagram to show that partitioning happens on high-level Relax IR before the LegalizeOps pass, and updating the C++ FFI example to use GetGlobal with an existence check instead of GetGlobalRequired.