Fix issue with reduce_vars

[L2jLiga]

Fixes #354
Fixes #294
Fixes #308

[L2jLiga]

@fabiosantoscode any updates?

[fabiosantoscode]

I'm very concerned about this: https://github.com/terser-js/terser/pull/363/files#diff-7a02108b11616cdd07d36b9a138a71cbR1290

undefined_variable++ is NaN, why is there even a zero there?

[fabiosantoscode]

Nevermind the NaN thing and the comment @L2jLiga, I'm missing the fact that this is not used only in the ++ operation.

This PR looks great. Let me investigate what could be the impact of this transformation to a 0 to ensure everything is going to be fine. Don't want to break anything anymore :D

[L2jLiga]

I think impact will be suboptimal result for some cases, so I suggest to create an issue to not forget to investigate/fix this

[fabiosantoscode]

@L2jLiga I've just had a long look. The impact isn't relevant enough that I care about it.

The 0; would only be output when the variable is an AST_Unary which is unused, which can really only happen if its parent is a single-expression statement. However, I don't want to increase the complexity of terser for a 2 byte reduction which would only take place if someone uses an unary operation on an unused variable, which is a pretty silly thing to do by hand.

Please remove the TODO comment and this PR can go in. Awesome work!

[L2jLiga]

TODO comment removed and PR rebased on top of master :)

[L2jLiga]

@fabiosantoscode let's land this

[fabiosantoscode]

Right. I forgot to merge. But I didn't want to ship because I'm working and if something breaks I can't handle it. Starting tomorrow I'm on vacation. Kind of. And I can ship the new version.