[App] Restore cpp_rtvm native inference tool on Relax VM and the tvm-ffi runtime

[Raman-Raje]

Summary

This PR brings back apps/cpp_rtvm, the standalone C++ tool (rtvm) for running TVM-compiled
models directly on a target device such as Android (OpenCL), without Python. It was removed in
#17660 along with the Relay C++ components. This version is rewritten for the current stack:

• Relax VM executables (vm_load_executable, vm_initialization, set_input,
  invoke_stateful, get_output) instead of the graph executor
• tvm::ffi APIs (ffi::Module::LoadFromFile, ffi::Function, Tensor)
• Per-backend runtime libraries (TVM_RUNTIME_BACKEND_LIBS), linked the same way as apps/cpp_rpc

Build with -DUSE_CPP_RTVM=ON (the option already exists in cmake/config.cmake).

Testing

• Built rtvm and libtvm_runner for CPU, GPU, and Android targets.
• Verified end-to-end inference on Android devices with the OpenCL and Vulkan backends.

[srkreddy1238]

LGTM

[Raman-Raje]

@tqchen

[tqchen]

given we are moving towards making the core more lightweight and now it is reasonably easy for agent to bring up the support, i feel it is nice to actually not having to maintain applications on the core repo

[Raman-Raje]

Thanks @tqchen, that makes sense given the direction of keeping the core lightweight.