Fix audio transcription deserialization when duration field is absent (#512)

The `duration` field on `AudioTranscriptionResponse` uses a custom
`deserialize_with`, which makes serde require the field to be present
in the JSON — even though it's `Option<f64>`. When the inference proxy
(vllm-proxy-rs) returns a response without a top-level `duration` field
(it nests it under `usage.seconds`), deserialization fails with
"missing field `duration`", causing a 502.

Fix: add `#[serde(default)]` so missing `duration` deserializes as `None`.

Author: Evrard-Nil

crates/inference_providers/src/models.rs

@@ -992,6 +992,7 @@ pub struct AudioTranscriptionResponse {
     pub text: String,
 
     /// Total audio duration in seconds (may be a string in some provider responses)
+    #[serde(default)]
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(deserialize_with = "deserialize_duration")]
     pub duration: Option<f64>,