N.B. install support for Python libaries/binaries was not working for the moment.

1) patchelf built from source to workaround creating-hole-thus-size-bloating
2) libtorch built with BUILD_CUSTOM_PROTOBUF=ON to avoid protobuf-doubly-linking
   issue

To avoid the bug in Apollo-simulator that complains:

```
File already exists in database: descriptor.proto
[google/protobuf/descriptor.cc:1359] CHECK failed: GeneratedDatabase()->Add(encoded_file_descriptor, size):
libc++abi.dylib: terminating with uncaught exception of type google::protobuf::FatalException: CHECK failed: GeneratedDatabase()->Add(encoded_file_descriptor, size):
```

1) APT provided version was new enough
2) To reduce image size

To avoid the patchelf-creating-holes-in-binaries-thus-causing-size-bloating problem

Docker: update {cyber,dev}.x86_64.stable images based on CUDA-11.1 to support Nvidia GPUs of Ampere architecture

Please note if you run `./apollo.sh build --config=cpu` on a GPU capable
machine, LD_LIBRARY_PATH may be wrong. We will revisit this in later PRs.

* Perception: resume traffic light module

1) cuda-11.1-x86_64
2) cuda-10.2-aarch64

1) libtorch-1.6.0 was built with magma support
2) Fix symlink issues of mkl and tbb PYPI packages