`CREATE EXTENSION btree_gin;` + `ops: raw("text_ops")` leads to endless migrations (and operator class is removed in Introspection)

(Originally reported by @uhyo in https://github.com/prisma/prisma/issues/16275#issuecomment-1402212976)

Currently, Prisma can not handle the operator classes added by extensions (see issue for support: https://github.com/prisma/prisma/issues/17516). This means to create an index using such an operator class in e.g. a Gin index, you have to use the raw("...") operator class escape hatch and then use migrate dev --create-only and then modify the migration file to make sure the extension is enabled:

generator client {
  provider = "prisma-client-js"
}

datasource db {
  provider = "postgresql"
  url      = env("DATABASE_URL")
}

model A {
  id   Int    @id @default(autoincrement())
  data String

  @@index([data(ops: raw("text_ops"))], type: Gin)
}

Then add the first line to the migration file before the correctly generated SQL:

CREATE EXTENSION btree_gin;
CREATE INDEX "A_data_idx" ON "A" USING GIN ("data" text_ops);

This migration can be applied with migrate deploy, but any future migrate dev run will again create a useless migration like this:

-- DropIndex
DROP INDEX "A_data_idx";

-- CreateIndex
CREATE INDEX "A_data_idx" ON "A" USING GIN ("data" text_ops);

(Note how the index definition is exactly the same as the original index definition that was already run)

Running Introspection will also remove the raw("text_ops") operator class from the original Prisma schema.

Reproduce the Prisma schema and SQL shown in the issue, then compare repeated migrate dev runs with migrate deploy and with introspection. Trace the migration diff and introspection paths that handle the Gin index and raw("text_ops") operator class. Done means the existing index is not recreated on later migrations and introspection preserves the operator class.