Make ConverterManager.getInstance() init thread-safe.

[cpovirk]

We encountered a TSAN error because 2 threads were racing to call
ConverterManager.getInstance(). My read is that the current code could
actually be unsafe, since one thread might see INSTANCE as non-null
before the other thread has initialized its fields.

Fortunately, this looks like a good case for the
initialization-on-demand holder
idiom.

(There would be additional thread-safety concerns if anyone were to try
to mutate the ConverterManager instance concurrently with usage. But
I'm not sure there's a solution to that short of spraying synchronized
around. That seems like probably overkill, given the possible
performance impact for what I'd hope is an uncommon use case.)

[jodastephen]

Looks like a good change and happy to merge, but unfortunately TestConverterManager needs fixing.

[cpovirk]

Sorry, I did not anticipate reflection in the tests :) I'll take care of it and then actually run tests, which apparently I had neglected to do.

[cpovirk]

The tests now pass.

I updated the assertion to look for the field in the new class. I now check that that class is private, but I figured that whether the field is private is mostly irrelevant in that case. Rather than throw away the field check entirely, I changed it to a check for final, since we're now able to use that.

But I'm happy to update it in some other way if you think that would be best.

[jodastephen]

Thanks!