[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

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
• Moving libtvm_runtime_cuda.so out of lib/ — CUDA unavailable, no crash

[gemini-code-assist]

Code Review

This pull request refactors the TVM runtime build system to split the monolithic runtime into a core library, per-backend dynamic shared objects (DSOs) for CUDA, Vulkan, OpenCL, Metal, ROCm, and Hexagon, and a separate libtvm_runtime_extra library for contrib and disco sources. Correspondingly, the Python loading logic was updated to dynamically discover and load these backend DSOs. Review feedback identified a missing compile definition for the cuDNN frontend and a package naming inconsistency in the library loading utility that could lead to resolution failures.

[tqchen]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors the TVM runtime build system to split backend-specific code (CUDA, Vulkan, OpenCL, Metal, ROCm, Hexagon) and various contrib modules into separate dynamic shared libraries, specifically introducing libtvm_runtime_extra. This change allows for a more modular runtime and dynamic loading of backends in Python. Feedback includes concerns regarding regressions for monolithic Hexagon DSP builds where Disco sources might be missing, the need to gate the creation of the extra runtime library for Hexagon, and improving the robustness of the CMake logic when no extra sources are present.

[gemini-code-assist]

The libtvm_runtime_extra assembly block should be gated to avoid creating a shared library when building for the Hexagon DSP (BUILD_FOR_HEXAGON). Hexagon DSP builds typically use static linking and a monolithic runtime, and the standard shared library mechanism used here may not be appropriate for the DSP environment.

# ---- libtvm_runtime_extra assembly ----
if(NOT BUILD_FOR_HEXAGON)

[tqchen]

Reopening from upstream branch to pick up Jenkins groovy changes.