Optimize field set for oneof case

[yanivshaked]

The current implementation initializes an empty List with n fields for a one-of case.
This means, that while one of of the elements in the list will actually be populated, the entire list is initialized taking n times the amount of memory than required.
The suggested change uses a Map instead of a List for the one-of case, reducing amount of memory drastically if one-of has many types to choose from (which is mostly the case).

[yanivshaked]

Updated files according to test results

[osa1]

The problem is we represent oneofs as message even though they have nothing in common.

We should have a separate value class for oneofs (similar to ProtobufEnum, which I wish we called PbEnum for consistency with PbList and PbMap) instead of optimizing message code (_FieldSet) to support this use case.

Message values are like product types (tuples, records, structs) and they're represented as GeneratedMessage.

Enum values are just integers with a mapping from the integers to names, and represented as ProtobufEnum.

Oneofs are sums (e.g. enhanced enums in Dart), and we don't have a way to represent sums efficiently currently. Instead we try to use GeneratedMessage which as demonstrated in this PR is not a good idea as it wastes O(n) space where n is the number of variants of the oneof.

[sigurdm]

Oneofs are sums (e.g. enhanced enums in Dart), and we don't have a way to represent sums efficiently currently. Instead we try to use GeneratedMessage which as demonstrated in this PR is not a good idea as it wastes O(n) space where n is the number of variants of the oneof.

I guess generating a class with a single dynamic field would be the easiest solution here without proper language support for sum-types.

We could generate the appropriate getters that will do the cast....

[osa1]

We can't make FieldSet._values dynamic (or any other type that will require polymorphism when accessing the elements) as that will make every operation slower.