Most strongly typed languages need each member of a union type to have an identifier. In Java we generate custom wrappers for unions, Rust needs names for enum variants (Rust unifies enums and tagged unions).

This PR updates the spec to change inline unions used in property declarations to a distinct type with a new @codegen_names (plural) annotation that names the union members.

class Foo {
  field: Bar | Baz
}

becomes

class Foo {
  field: SomeBarBaz
}

/** @codegen_names bar, baz */
type SomeBarBaz = Bar | Baz

Some inline unions are left in the spec, that represent the leniency rules that ES implements. These are notably SomeType | SomeType[] (read a single object as a 1-sized array) and strings that accept any primitive type.

These leniency rules are currently implemented in the Java generator and will later be formalized as new model validation rules that will report inline unions that require a separate type definition.

Along with refactoring enumerations to enum types, this PR also fixes a number of definitions that were ambiguous or actually wrong. API validation has been run on most of these changes to verify that there were no regressions.

Validation of the search API has 3 errors less, and the remaing ones are because of the use of numbers for ids (they're not quoted in the YAML tests) and experimental bucket_correlation and bucket_count_ks_test aggregations that have not been represented in the spec.