[BUILD] Modularize device runtime into per-backend DSOs

[tqchen]

Summary

Modularize libtvm_runtime into per-backend shared libraries (libtvm_runtime_cuda, libtvm_runtime_vulkan, etc.) and libtvm_runtime_extra for contrib/disco modules.

• Each backend can be built independently (USE_CUDA=ON produces only libtvm_runtime.so + libtvm_runtime_cuda.so)
• During Python import, available backend DSOs are discovered and loaded automatically, with missing backends silently skipped
• libtvm_runtime_extra is always produced if it has sources (disco, contrib modules)
• No user-visible config flag changes — existing USE_CUDA, USE_VULKAN, etc. continue to work
• Each contrib module builds as its own OBJECT lib, linked via target_link_libraries (modern CMake pattern)
• CI artifact stashing updated to include new DSOs

Test plan

• Build with USE_CUDA=ON — verify libtvm_runtime.so, libtvm_runtime_cuda.so, libtvm_runtime_extra.so produced
• python -c "import tvm; print(tvm.cuda(0).exist)" prints True
• Runtime GPU tests pass
• Windows build passes (MSVC dllexport/dllimport)
• cuBLAS/cuDNN JSON runtime functions registered via libtvm_runtime_extra

[gemini-code-assist]

Code Review

This pull request refactors the TVM build system to transition from a monolithic runtime to a modular structure using multiple shared libraries, including per-backend DSOs and a dedicated library for optional components. Key changes include updates to CMake modules for device runtimes, refined Python library loading logic, and adjusted CI artifact handling. Review feedback identifies a critical issue regarding duplicate symbol definitions in the CLML module due to redundant source inclusion and recommends more specific exception handling in the CI S3 script to prevent masking infrastructure failures.