Fix #3348: use a new typeState in inferView

[liufengyun]

Fix #3348

inferImplicit has a useful invariant that the constraint doesn't
change in the case of implicit resolution failure. However, this
invariant may not hold if completions are triggered during implicit resolution.
By create a new typeState for inferView, we ensure that the useful
invariant holds in such cases.

[liufengyun]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/3354/ to see the changes.

Benchmarks is based on merging with master (174444d)

[allanrenucci]

You forgot to add the test case

[odersky]

I am not sure anymore whether we should undo completions when an implicit search fails. We have at least three choices here:

1. Simply remove the assertion, which means we accept that a failed implicit search can add to the constraint via completions.
2. Create a new TyperState and commit only on success, as done in this PR. But note that this will also forget all error messages that would be emitted for completed symbols
3. Reset the constraint only, the same way it is done in isSubType.

I think we need to dig deeper to decide which of the three is the right strategy.

[liufengyun]

My feeling is that the assertion is a useful invariant to guard against other mistakes, thus I would make choice 1 as last resort.

For 2nd solution, actually the errors that emitted during completion are reported, as the following example shows:

class Test {
  import Test.test
  "Hello".toto     // error
}

object Test {
  def test = {
    abc                  // <== error reported
    implicitly[collection.generic.CanBuildFrom[List[Int], Int, List[Int]]]
  }
}

Or there are may be other cases that I'm ignorant of.

The 3rd option sounds good, I'll have a look.

[odersky]

LGTM