Support mixed text and raw-token rollout content in Swift template

[kerbeans]

Summary

Support mixed rollout response content in Template._swift_encode().

Rollout responses may contain multiple content components, including a text segment followed by a raw-token segment, for example:

[
    "<think>...</think><tool_call>...</tool_call>",
    {
        "loss_scale": [...],
        "token_ids": [...]
    }
]

The current Swift template encoding path assumes that a list whose final item is not a string can be treated directly as List[int]. For mixed content such as List[str, dict], this passes the entire list to tokenizer.decode(), causing a type error.

This change recognizes a trailing raw-token component and uses its token_ids only for suffix inspection. The original response representation is preserved for template encoding, so raw-token rollout semantics are unchanged.

Tests

• tests/general/test_template_mixed_rollout.py: 5 passed
• Existing exact-token rollout tests: 33 passed; one unrelated existing test requires a missing trainer_name fixture