[Draft] AST based cache

[martinbonnin]

Add an alternate cache that reads the selections from operation.document instead of using codegen and CompiledField.

This trades codegen for runtime parsing. Current thoughts:

1. addTypename needs to be "always" if we don't want to codegen the full type definitions. This is a sensible default when using the cache anyways because a given object type might be accessed from an interface.
2. because GQLField has no type (again because we don't want to codegen the full type definitions), it's impossible to tell if a JSON object is scalar or not, which means scalar encoded directly in the JSON will now create intermediate records. Not 100% sure how much of a problem that is.
3. CompiledField is a public API in cacheKeyGenerator. If some people rely on rawType() this is going to be a problem. On the other hand I think using rawType() is a mistake in most cases as what's included in the cache key is __typename instead.
4. TODO: put back keyFields and paginationArgs support.
5. TODO: migrate data builders to use GQLField instead of CompiledField
6. TODO: add macrobenchmarks

Current microbenchmarks (query1 is using CompiledField, query2 is using AST):

26,353,732   ns      183675 allocs    trace   .query1Sql
31,018,269   ns      261107 allocs    trace   .query2Sql
4,107,074   ns      115242 allocs    trace   .query1Memory
6,154,897   ns      185074 allocs    trace   .query2Memory

While it's slower, microbenchmarks are not very representative because it parses the document every loop while the CompiledField is built only once during classloading and kept in memory after that.

[netlify]

✅ Deploy Preview for apollo-android-docs canceled.

Name	Link
🔨 Latest commit	75c8d52
🔍 Latest deploy log	https://app.netlify.com/sites/apollo-android-docs/deploys/64aefc6046f3bb0008c3c846