Docker: Add IO::Async::Loop::Epoll for IPerl

This changes the default event loop used inside of IPerl from
`IO::Async::Loop::Poll` to `IO::Async::Loop::Epoll`.

This is needed to address a SIGSEGV that would consistently occur for
GPU containers.  This SIGSEGV would only occur when running the code
using IPerl inside of a Docker container. To clarify,

  - running the code as a regular script worked inside of Docker,
  - as did running the code as both a regular script and inside of IPerl
    on the host.

It's not clear why the SIGSEGV would occur only when running inside of a
Docker container, but by looking at the core dump and using a debugging
build of TensorFlow, it seems to always happen when `libtensorflow`
would call out to a subprocess such as when it would try to use `ptxas`
from the CUDA toolkit (across multiple versions of the CUDA toolkit).
The core dumps would usually contain a stack backtrace of

  - `tsl::SubProcess::WaitInternal` ;
  - `tsl::SubProcess::Communicate` ;
  - `stream_executor::GetPtxasVersionString` .

When running under a debugger, stepping line-by-line over the calls to
the subprocess would sometimes make the SIGSEGV go away, but this was
not consistent.

My theory is that the calls to the subprocess also use the `poll(2)`
syscall and this is causing some kind of interaction between the two
event loops. Switching to `epoll(7)` means this interaction goes away.

Author: zmughal

docker/Dockerfile

@@ -54,7 +54,7 @@ RUN bash -c 'cd /tf \
 	'
 
 # Install Jupyter kernel
-RUN plx cpm install -L /perl5 Devel::IPerl
+RUN plx cpm install -L /perl5 Devel::IPerl IO::Async::Loop::Epoll
 
 # Install libtensorflow (with Alien deps to /perl5 and actual Alien dist to
 # /perl5-libtf)