Structured config created from subclass of dataclass does not have correct initial values

[Jasha10]

Describe the bug
When an omegaconf structured config is created from a subclass of a dataclass, the config is not initialized with the default values from the subclass; instead the default values from the superclass are used.

To Reproduce

from dataclasses import dataclass, field
from typing import Any, Dict, List

from omegaconf import MISSING, OmegaConf


@dataclass
class ParentConfig:
    model_depth: int = MISSING
    model_list: List[int] = field(default_factory=lambda: [5, 6])
    model_list_2: List[int] = MISSING
    model_dict: Dict[str, Any] = MISSING


@dataclass
class ChildConfig(ParentConfig):
    model_depth: int = 5
    model_list_2: List[int] = field(default_factory=lambda: [1, 2, 3])
    model_dict: Dict[str, Any] = field(default_factory=lambda: {"a": 5, "b": 6})


cfg = OmegaConf.structured(ChildConfig)
print(OmegaConf.to_yaml(cfg))

# Actual Output:

model_depth: 5
model_list:
- 5
- 6
model_list_2: ???
model_dict: ???

# Expected Output:

model_depth: 5
model_list:
- 5
- 6
model_list_2:
- 1
- 2
- 3
model_dict:
  a: 5
  b: 6

This is a dataclass issue; when using attr classes instead of dataclasses, there is no bug.

Additional context

• Based on this issue: [Bug] Override MISSING value in child config facebookresearch/hydra#1881