grpc-js: Hold a reference to transport in SubchannelCall

[murgatroid99]

#2318 (comment) reported a regression in that previously fixed issue. I suspect that the problem is that once a subchannel stops using a transport, nothing has a reference to that transport, which makes it eligible to be garbage collected. To avoid this, the subchannel call now holds a reference to the transport, so it stays alive as long as at least one active call is using it.

I also removed the disconnectListener because it became unused in the previous refactor.

[gnossen]

Is there a way to reproduce this issue in a unit test?

[murgatroid99]

The problem with bugs like this is that the test framework code seems to change the behavior. I do have a set of local interop tests that run the client in a separate process in order to catch some bugs like this, but they don't catch this one and I don't know why. In this specific case, if I'm right and it's tied to garbage collection, that makes it even more unpredictable.

[yamen]

Any open "ref" in my scripts would cause the bug not to appear (presumably because GC isn't triggered at the same time). But it appears 100% consistently in scripts that are very simple and guarantee no other open refs.