fix(azure): omit temperature for unrecognised deployment names

[Aftabbs]

Summary

Fixes #3015.

Azure users commonly pass a deployment name (e.g. gpt5-prod-eu, my-gpt4-deployment) as the model argument instead of the canonical model name. When that string is absent from the OPENAI_MODELS_DATA catalogue, ModelDataRegistry.get() falls back to DeepEvalModelData() whose supports_temperature defaults to True, so temperature=0.0 is forwarded to the API. Newer Azure-hosted models (GPT-5 family and any future model) reject the temperature parameter and return a 400 error.

Root cause

# azure_model.py (before)
if self.model_data and self.model_data.supports_temperature is False:
    temperature = None

The guard only fires when supports_temperature is explicitly False (i.e. a known reasoning model). An unrecognised deployment name returns the default DeepEvalModelData() with supports_temperature=True, so the guard never fires and temperature is sent regardless.

Fix

# azure_model.py (after)
if self.model_data.supports_temperature is False or model not in OPENAI_MODELS_DATA:
    temperature = None

Two cases now correctly suppress temperature:

1. Known reasoning models – supports_temperature=False explicitly (o-series).
2. Unknown deployment names – not present in the catalogue; we cannot confirm support, so we omit temperature defensively.

Known standard models (gpt-4o, gpt-4-turbo, etc.) remain unaffected because they are in the catalogue.

Test plan

• TestAzureModelTemperature::test_unknown_deployment_name_temperature_is_none – unrecognised model name → temperature is None by default.
• TestAzureModelTemperature::test_unknown_deployment_name_explicit_temperature_also_none – user-supplied temperature still dropped for unrecognised names.
• All 4 pre-existing temperature tests continue to pass (no regression for gpt-4o or o3-mini).

tests/test_core/test_models/test_azure_model.py::TestAzureModelTemperature - 6 passed
tests/test_core/test_models/test_azure_model.py - 32 passed

[vercel]

@Aftabbs is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.