Fix: honor chat_default as fallback chat model for subscribed (self-hosted) users

[SnowingFox]

Problem

On self-hosted Khoj (billing disabled), every user is treated as subscribed by ais_user_subscribed(). In the async chat-model-selection path, ConversationAdapters.aget_default_chat_model() only ever returned chat_advanced for subscribed users and never fell back to chat_default. This made chat_default dead config on self-hosted installs and forced self-hosters to use chat_advanced for every LLM call (intent classification, code generation, title generation, chat responses), regardless of the sidebar model picker.

Root cause

In src/khoj/database/adapters/__init__.py, the subscribed branch of aget_default_chat_model() checks think_paid_fast, then think_paid_deep, then chat_advanced — but, unlike the non-subscribed branch and the synchronous get_default_chat_model(), it never falls back to chat_default. Because ais_user_subscribed() returns True whenever state.billing_enabled is False, self-hosted users always take the subscribed branch, so the server admin's chat_default model is never selected.

Fix

Add the missing chat_default fallback to the subscribed branch of aget_default_chat_model(), mirroring the non-subscribed branch and the existing synchronous get_default_chat_model(). When chat_advanced (and the think-paid slots) are not configured, the server admin's chat_default model is now used. When chat_advanced is set, it still takes priority, preserving existing behavior.

Test

New tests/test_chat_model_selection.py:

• test_aget_default_chat_model_honors_chat_default_when_chat_advanced_unset — with a self-hosted (billing disabled) user and only chat_default configured, aget_default_chat_model returns chat_default. Fails before the fix (returns the first-created chat model instead) and passes after.
• test_aget_default_chat_model_prefers_chat_advanced_when_set — locks in that chat_advanced still wins when it is configured.

Fixes #1282