Binding SPI is a little strange when it comes to Provider bindings

```
@Test public void testProviderSpi() {
Module providesModule = new AbstractModule() {
@Override protected void configure() {
bind(Key.get(Object.class)).toInstance(new Object());
}
};

Injector rootInjector = Guice.createInjector(providesModule);
Set> keys = rootInjector.getAllBindings().keySet();
rootInjector.getExistingBinding(new Key>() {});
Set> keys2 = rootInjector.getAllBindings().keySet();
assertThat(Sets.difference(keys2, keys)).isEmpty();
}
```

This test fails with:
java.lang.AssertionError: Not true that <[Key[type=com.google.inject.Provider, annotation=[none]]]> is empty

This is because of special whitelisting that getExistingBinding does when it comes to allocating provider bindings.

> There's some weirdness internally in that Guice is perfectly happy caching bindings for Provider> if someone happens to call getInstance(Key>) too. Seems like it'd be better to handle Providers by using Binding.getProvider(). But those structures generally predate my involvement, and I've never had time or a reason to dig through it and/or clean things up.
> @sameb

So maybe there is some potential for improvement on the representation of provider bindings.