Previously on downgrade we could pick the wrong version.

This will correct the issue going forward, but we have no way of fixing
history

Fixes: #171

In single_threaded mode MiniRacer will behave far more reasonably if platform
is initialized in the master process.

Sadly this is only a partial solution since we will hang at:

```
#0  0x00007fe6f037e34d in pthread_cond_broadcast@@GLIBC_2.3.2 () from /usr/lib/libpthread.so.0
#1  0x00007fe6ebc15c05 in v8::platform::DelayedTaskQueue::Terminate() ()
   from /home/sam/Source/mini_racer/lib/mini_racer_extension.so
#2  0x00007fe6ebc14dfd in v8::platform::DefaultWorkerThreadsTaskRunner::Terminate() ()
   from /home/sam/Source/mini_racer/lib/mini_racer_extension.so
#3  0x00007fe6ebc13bd6 in v8::platform::DefaultPlatform::~DefaultPlatform() ()
   from /home/sam/Source/mini_racer/lib/mini_racer_extension.so
#4  0x00007fe6ebc13cde in v8::platform::DefaultPlatform::~DefaultPlatform() ()
   from /home/sam/Source/mini_racer/lib/mini_racer_extension.so
#5  0x00007fe6effb0db7 in __run_exit_handlers () from /usr/lib/libc.so.6
```

Likely some more v8 changes are needed to mitigate the heavily forked use
case.

Adds a custom loader for MiniRacer that isolates all symbols, this allows loading multiple versions of v8 into a single process.

Our exit cleanup could hand after forking, this ensures that post fork process can still exit cleanly.

This causes issues on some Ruby installs, e.g AWS Linux
Keeping it around for diagnostics/future configurable

Required for recent-ish C++ required by current V8