litellm: trace decisions media inputs and patch systemone/asystemone #970

Upstream (litellm 1.106.0): litellm.decisions() / adecisions() gained audio and videos parameters, and state= calls now forward images and assets. decisions(state=...) is deprecated in favour of litellm.systemone() / asystemone(), which already existed in 1.105.

Today:

• _decisions_span_payload (py/src/braintrust/integrations/litellm/tracing.py:356) builds the span input from state/input and questions only. The new media inputs never reach the span.
• systemone isn't patched, so calls to it produce no span. Since decisions(state=...) is deprecated in its favour, users will increasingly move to an untraced call.

Change:

• Add the media inputs to the span input using the attachment conversion feat(integrations): support multimodal decision spans #941 added in py/src/braintrust/integrations/question.py.
• Add patchers for systemone / asystemone that reuse the decision-span path.

Effort: Small to medium. Needs new cassettes for a media decision and a systemone call.