Start moving toward a callback model for code-completion type-checking

[benlangmuir]

... to reduce and someday eliminate all the places we do double
typechecking. The goal is to setup any 'interesting' expressions ahead
of time so that if we type-check them at any point - e.g. during
checking the context - we don't lose the results and try to check them
again.

For now, this is limited to cases where the original expression is not
replaced by pre-checking or CSGen (i.e. we don't handle
UnresolvedDeclRef through this mechanism yet), and we're only using it
for the "parsed expression", which is either the base of a dot
completion, or the name in a function-call or initializer call
expression. The intent is to move unresolved member completion to the
same model in the future.

Currently, we use the results from the callbacks for 60% of these
completions, and of the 40% we miss, about 80% are simply
UnresolvedDeclRef or similarly obvious cases.

Progress toward rdar://21466394
Incidentally fixes rdar://20962472

[benlangmuir]

@swift-ci Please smoke test

[benlangmuir]

@swift-ci Please ASAN test

[benlangmuir]

Asan bot timed out, but it got through all the tests I was interested in anyway.

[benlangmuir]

@DougGregor and/or @rudkx could you glance at the type-checker side of my changes?

[rudkx]

@benlangmuir Is this PR still worth keeping open?

[I'm going through all the open PRs where I'm @-mentioned to see if there is anything I need to follow-up on.]

[benlangmuir]

Closing. I have no immediate plans to work on this and it's probably the wrong direction anyway.