Get Llm Proxy Requests For Call

One call's requests in order, with the computed request_seq.

request_seq orders reading; it is not a join key against vapi_turn_latency.turn_index. Vapi retries a failed turn 3x, a tool-calling turn is one Vapi turn but 2+ requests, and a model-generated first message is a request with no user turn at all. Align the two positionally with a FULL OUTER JOIN so divergence shows up as one-sided NULLs instead of being silently dropped.