Member lookup macro expansion performance

[DougGregor]

When performing name lookup into a type (and its extensions), we may need to expand macros. While we are careful to expand macros that we don't need to expand, we were still looping over all of the extensions of a type (and all of the members of those extensions!) looking for places where macros could have been used. In an example with 6,000 extensions of the same type, generated like this:

let upperLimit = 6000

print("""
  struct MyType {
    func f0() { }
  }
  """)

for i in 1..<upperLimit {
  print("""
    extension MyType {
      func f\(i)() {
        f\(i-1)()
      }
    }  
    """)
}

The checking for potential macro expansions regressed type checker performance by more than 25%, and adding on more extensions compounded the problem nonlinearly.

Introduce mechanisms to ensure that we only check what kinds of macro expansions each extension introduces once, capturing all of the potential macro-introduced names and caching it via the evaluator. Then, keep a list of only the extensions that can produce macro expansions and loop over that---not the whole list of extensions---when we need to check whether a given name might be produced by macros. This returns the example's type-checking time back to what it was

Fixes rdar://109543968, eliminating the regression.

[DougGregor]

@swift-ci please smoke test

[DougGregor]

@swift-ci please smoke test

[DougGregor]

@swift-ci please test compiler performance