Change the conversion rule to favor any number of widenings (CGFloat -> Double)
over even a single narrowing conversion and if there is no way to avoid narrowing
(due to contextual requirements) attempt it as late as possible (the deeper in
the AST that conversion is located the higher its score).
This is a generally better rule when it comes to rounding and information
loss that would result from narrowing conversions.
This wasn't a problem before since locator wasn't really used by
`ExprRewritter:coerceToType` but with Double<->CGFloat conversion
it needs the locator to be anchored at a rewritten expression instead
of the original one to form a correct implicit initializer call.
