Add tests cases for, and fix, #513

[mayoff]

This patch addresses both leaks discussed in #513.

First leak

The changes in d1d6ca4 made several Geometry2DTests fail.

The actual unbalanced retain causing the leak in VariantRepresentable.swift:

extension ContentVariantRepresentable {
    public init? (_ variant: Variant) {
        guard Self.godotType == variant.gtype else { return nil }

        var content = Self.zero
        withUnsafeMutablePointer(to: &content) { ptr in
            variant.toType(Self.godotType, dest: ptr) // <-- THIS IS THE UNBALANCED RETAIN
        }

        self.init(content: content)
    }
}

The variant.toType function calls into Godot to copy out the built-in type stored in variant. If that built-in type has an internal reference count, Godot increments it. Then, the call to self.init(content: content) increments the count again. The retain performed by variant.toType is never balanced.

This patch fixes the bug by generating a new init(alreadyOwnedContent:) initializer for each Godot builtin class type. This needs to be on the builtin wrappers (like GDictionary, GArray, etc.), rather than on Variant which is where d1d6ca4 put it.

This patch also adds a test case to check for the leak by looking at Godot's MEMORY_STATIC performance counter, which is only enabled if Godot was built with DEBUG_ENABLED.

Second leak

This patch adds a test for the second leak described in #513, and fixes the leak.

Without this patch, the leak happens because Variant.asObject has an unneeded call to rc.reference() which increments the RefCounted object's reference count. As far as I can tell, nothing balances this increment.

Variant.asObject calls lookupObject(nativeHandle:), which returns a Swift wrapper for an object whose reference count has already been incremented (if the object is RefCounted). The Swift wrapper balances that increment with a decrement in its deinit.

[migueldeicaza]

You are absolutely right about these observations, I ran into these while cooking the tests, and noticed the 2D tests fail and spent some quality time trying to figure it out, when I saw that I decided to check on the bug again and saw your analysis.

Love your patch, and the additional insight on the differences on those built-in types.