Python: Harden Python checkpoint persistence defaults

[moonbox3]

Persisted checkpoint handling in the Python SDK should be hardened so the default experience is more restrictive and safer by default.

Today, checkpoint persistence is optimized for broad object round-tripping and ease of use. Before GA, we should tighten the default behavior for persisted checkpoints and reduce the default type surface area that can be restored.

Proposed direction:

• Narrow the default set of types that round-trip from persisted checkpoints.
• Include a small built-in set of framework-approved types by default, such as primitives and other common value types.
• Add an explicit allow-list for additional checkpoint-safe types.
• Update docs, tests, and samples to reflect the safer default posture.

Example API shapes we could evaluate:

storage = FileCheckpointStorage("/tmp/checkpoints")

storage = FileCheckpointStorage(
    "/tmp/checkpoints",
    allowed_checkpoint_types=[
        "my_app.models.SafeState",
    ],
)

Built-in supported types could include framework-approved value types such as primitives plus selected common types like datetime.datetime and uuid.UUID.

Acceptance criteria:

• Default persisted checkpoint loading is more restrictive than it is today.
• Common framework-owned checkpoint scenarios continue to work without extra user configuration.
• Developers have a documented extension point for additional supported types.
• Docs and samples reflect the updated default guidance.

Open questions:

• What is the minimum built-in type surface area needed to support common workflow scenarios?
• What upgrade guidance do we want for existing persisted checkpoints?