Fix dict filter by removing items, Fix tests for v0.3 compat

[ranjanan]

Fix the function filter!() by removing items(d)

[ranjanan]

This should fix #226

[ranjanan]

@tkelman Could you please help with code review?

[ranjanan]

The Appveyor build fails because of this. There needs to be an @compat there for all the UInt8's. But I'm curious about why Appveyor checks against v0.3 while Travis checks against v0.4

[tkelman]

The .travis.yml file here is still using the PPA for installing Julia, which does not make it very easy to install 0.3 any more. You can't use multiple language: settings, and it's a little more effort to install specific versions of Python than it is specific versions of Julia. The Travis file could be reformatted to download the generic Linux binaries (and osx dmg installers if osx is running) instead of the PPA to restore testing against 0.3.

[ranjanan]

I see. Maybe I should modify Travis to do pull the v0.3 binaries and send a separate PR. Also, the above commit fixes #228 . Actually, I belive the problem is that this test fails on 32 bit Windows. Interestingly, this is a regression between the latest release and the current master. I used a Windows VM and Pkg.test() worked on the last release but breaks on master.

[stevengj]

Doing for (k,v) in d uses the dictionary iterator, which calls PyDict_Next. The documentation for the latter function explicitly says "The dictionary p should not be mutated during iteration"

Hence, I don't think this change is a good idea; there must be a more reliable way to fix it.

[stevengj]

Two options that I can see:

• Replace items(d) (which must have gotten removed from Julia at some point) by e.g. collect(Tuple{PyObject,PyObject}, d). This unfortunately makes a copy of the data.
• Do for (k,v) in d or similar, but instead of deleting the keys right away, push! them to a temporary array of keys to be deleted. This is a bit more memory-efficient for large dictionaries.

[stevengj]

Fixed by 73c3db7