AttributeError in Telemetry._cache_buckets for providers without prompt caching

Any provider whose usage response includes prompt_tokens_details without cache fields crashes profile validation and token recording with:

'PromptTokensDetailsWrapper' object has no attribute 'cache_creation_tokens'

Reproduced with MiniMax (Custom Model openai/MiniMax-M3, also the native minimax/MiniMax-M3 provider, Base URL https://api.minimax.io/v1, openhands-sdk 1.49.1), and other reporters hit it with different providers and models, including GLM-5.3 (see #5099 and the comments below). Also verified working with ZAI glm-5.3-flash.

Root cause: Telemetry._cache_buckets in openhands-sdk/openhands/sdk/llm/utils/telemetry.py reads details.cache_creation_tokens directly, guarded only by "cache_creation_tokens" in details.model_fields_set. LiteLLM's PromptTokensDetailsWrapper.__init__ deletes unset optional fields, while its __setattr__ mirroring keeps the name in model_fields_set. So the guard passes and the read raises AttributeError.

Actual Behavior

Saving the profile fails validation with the AttributeError above. Reproduced against the installed SDK with python:

from litellm.types.utils import Usage, PromptTokensDetailsWrapper
from openhands.sdk.llm.utils.telemetry import Telemetry

usage = Usage(
    prompt_tokens=100,
    completion_tokens=50,
    total_tokens=150,
    prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=0),
)
Telemetry._cache_buckets(usage)

model_fields_set contains cache_creation_tokens even though the attribute was deleted, so the guard passes and the direct read raises.

Acceptance Criteria

• Saving an LLM profile for a provider without prompt caching passes validation instead of raising AttributeError
• The cache buckets default to 0 when the provider returns no cache fields

Start in openhands-sdk/openhands/sdk/llm/utils/telemetry.py at Telemetry._cache_buckets, then run the Python reproduction using Usage and PromptTokensDetailsWrapper shown in the issue. Confirm that profile validation and token recording no longer raise AttributeError and that cache buckets default to 0 when cache fields are absent.