Avoid using magic string values for JsonNull/DbNull

[matthijs-relay]

Problem

Since Prisma 3.0, the input for filtering and updating nullable JSON fields no longer accepts null, but instead the Prisma.JsonNull and Prisma.DbNull values to better distinguish between the two types of nulls.

I can see that this solves a real problem, but the current implementation has a significant drawback: JsonNull and DbNull are implemented as magic string values. This means their types overlap with regular string data that can be stored in a JSON field. Some problems this could cause:

• Filtering/storing Prisma.DbNull on a non-nullable JSON field will type-check, but produce a JSON string value of "DbNull" instead of an actual NULL, which may lead to confusing bugs
• User-provided strings that get stored as a top-level value in a JSON field need to be sanitized to make sure they're not "JsonNull" or "DbNull", or users might be able to inject either kind of null value and cause an unexpected state

Suggested solution

A possible alternative approach would be making JsonNull and DbNull Symbols or some kind of class instance, so that they don't overlap with the JsonValue type

[aqrln]

This issue is now fixed, and the change will be released in Prisma 4.0.0.

DbNull, JsonNull and AnyNull will now be singleton instances of their eponymous classes. We eventually settled on this design after making them symbols first because it allows for better type errors, and thus better DX (but you may still see "symbols" being referenced in some public GitHub issues used for planning and orchestrating this change).