Reliability Improvements on Transport Tier

[roamm]

Firstly I'd like to say I don't run into any obvious bugs without these changes.

Because Socket has some implicit assumptions like:

1. never call this.transport.close() more than once;
2. never rely on 'close' event from the underlying transport for a forced close;
3. the polling transport can be in an 'opening' state even Socket has closed;
4. callback of transport.close() is not reentrant;
5. 'close' event (not triggered in current implementation) from the old transport will break the logic while upgrading;

These assumptions make the code kinda fragile so I took some time trying to fix it and hope this would help.

[rauchg]

@roamm can we make more focused pull requests? Like one for idempotent close, that could be a PR by itself. Will make reviewing easier

[rauchg]

Thanks!

[plievone]

@guille, are there important failing test cases here? roamm@aa0209f