[ConstraintSystem] Don't warn about some redundant coercions

[xedin]

Let's not warn about possible redundant coercions if the
"coerced type" is not known upfront, because that would
mean that solver would infer it from coercion.

Resolves: rdar://problem/56608085

[xedin]

/cc @LucianoPAlmeida. Unfortunately looks like some warnings are too eager. This is showing up in validation-test/Sema/type_checker_perf/fast/rdar17077404.swift which is only run in --no-assertions build at the moment.

[xedin]

@swift-ci please smoke test

[xedin]

If this breaks something else I'll have to revert your original PR :/

[xedin]

Looks like this breaks cases like let _: (Float) -> Void = { $0 as Float } :(

[LucianoPAlmeida]

Right, that makes sense.
I see you've opened a PR to revert, thank you @xedin :)
Is that the only case broken? I think I can open a PR fixing that right now so we don't have to revert everything. But also can reopen the PR with the fixes 👍

[LucianoPAlmeida]

@xedin Just took continuation of this in another PR, hope you don't mind :)
Also, which should be the test we run on the CI to make sure everything is ok?