[Bug] USE_LIBTORCH is disabled in all builds

[leandron]

trafficstars

On my journey in enabling Integration tests in AArch64, I stumbled upon ctypes.tests.python.contrib.test_libtorch_ops.test_backend (see full error here), which crashes running integration tests on machines with torch installed.

The reason for that, is because TVM was not compiled with USE_LIBTORCH ON, and therefore the relevant sources won't be compiled into the output TVM libs. I am opening this as a tracking issue to enable USE_LIBTORCH, so that corresponding test(s) don't get stale.

This is also related to #12529, in which we probably want to enable Integration tests in CI_GPU.

In my view we need to:

• [ ] Document USE_LIBTORCH in our example config.cmake
• [ ] Activate USE_LIBTORCH in the build config script (the AArch64 seems the best positioned as of now)