Improve dataclass serialization

[andystaples]

Fixes two bugs with dataclass auto-serialization:
Bug 1 — Nested dataclasses lose the AUTO_SERIALIZED marker:
In shared.py, dataclasses.asdict(obj) was used, which recursively converts all nested dataclass fields into plain dicts. Only the outermost dict received the AUTO_SERIALIZED marker, so inner dataclass objects deserialized as plain dict instead of SimpleNamespace.

Fix: Replaced dataclasses.asdict(obj) with a shallow field extraction. This leaves nested dataclass values as-is, so the encoder calls default() on each one individually, giving each its own AUTO_SERIALIZED marker.

Bug 2 — SimpleNamespace mutation:
vars(obj) returns the object's actual __dict__, not a copy. Adding d[AUTO_SERIALIZED] = True permanently mutated the original SimpleNamespace by injecting an __durabletask_autoobject__ attribute.

Fix: Changed vars(obj) to dict(vars(obj)) to work on a copy.