struct field dependencies are overly conservative

[nikomatsakis]

The test cases I am adding in an upcoming PR show that for some reason we consider functions as dirty when any struct changes, even if the fn doesn't use that struct.

[nikomatsakis]

(Look for the FIXMEs.)

[nikomatsakis]

Heh, ok, I figured out why this is. The problem is that we use one dep-node for "trait selection on a given Trait". This is awfully coarse-grained, but I thought we could get away with it. Turns out, not such a good idea with the Sized trait.

[michaelwoerister]

:D

[nikomatsakis]

(got a fix locally, will prep PR soon)