`OverlappingFieldsCanBeMerged` would infinite loop when passed something like
```graphql
fragment A on User {
name
...A
}
```
It's not `OverlappingFieldsCanBeMerged`'s responsibility to detect that validation error, but we still would ideally avoid infinite looping.
This detects that case, and pretends that the infinite spread wasn't there for the purposes of this validation step.
Also, by memoizing and checking for self-references this removes duplicate reports.
ref: graphql/graphql-js#1111
This improves the overlapping fields validation performance and improves error reporting quality by separating the concepts of checking fields "within" a single collection of fields from checking fields "between" two different collections of fields. This ensures for deeply overlapping fields that nested fields are not checked against each other repeatedly. Extending this concept further, fragment spreads are no longer expanded inline before looking for conflicts, instead the fields within a fragment are compared to the fields with the selection set which contained the referencing fragment spread.
e.g.
```graphql
{
same: a
same: b
...X
}
fragment X on T {
same: c
same: d
}
```
In the above example, the initial query body is checked "within" so `a` is compared to `b`. Also, the fragment `X` is checked "within" so `c` is compared to `d`. Because of the fragment spread, the query body and fragment `X` are checked "between" so that `a` and `b` are each compared to `c` and `d`. In this trivial example, no fewer checks are performed, but in the case where fragments are referenced multiple times, this reduces the overall number of checks (regardless of memoization).
**BREAKING**: This can change the order of fields reported when a conflict arises when fragment spreads are involved. If you are checking the precise output of errors (e.g. for unit tests), you may find existing errors change from `"a" and "c" are different fields` to `"c" and "a" are different fields`.
From a perf point of view, this is fairly minor as the memoization "PairSet" was already keeping these repeated checks from consuming time, however this will reduce the number of memoized hits because of the algorithm improvement.
From an error reporting point of view, this reports nearest-common-ancestor issues when found in a fragment that comes later in the validation process. I've added a test which fails with the existing impl and now passes, as well as changed a comment.
This also fixes an error where validation issues could be missed because of an over-eager memoization. I've also modified the `PairSet` to be aware of both forms of memoization, also represented by a previously failing test.
ref: graphql/graphql-js#386
* Adds support for resolving union/interface types when using a generated schema
* Move resolveType __typename checking into defaultResolveType
* Clean up existing tests and improve error messages
ref: graphql/graphql-js#947
# Conflicts:
# src/Utils/BuildSchema.php
# tests/Utils/BuildSchemaTest.php
