feat: Update Qwen3 to Qwen3.5

[r-dh]

Summary

• Upgrade default local LLM from Qwen 3 (8B/4B) to Qwen 3.5 (9B/4B)
• Bump llama-cpp-python optional dep from >=0.3.9 to >=0.3.16
• Fix division by zero in _limit_chunkspans when context budget is exhausted

Blocker

Depends on abetlen/llama-cpp-python#2133, which adds Qwen 3.5 GDN (Gated Delta Network) support to llama-cpp-python. That PR also fixes a prefix-caching bug affecting all hybrid architecture models. Until it is merged and released as >=0.3.16, this PR cannot be merged.

Known issue

test_self_query fails: Qwen 3.5 returns {'topic': ['Physics']} instead of {} for an off-topic query ("What is the price of a Bugatti Chiron?"). The model applies a metadata filter even when the query is unrelated to the dataset. This is a behavioral regression compared to Qwen 3 and should be addressed separately, either by tuning the self-query prompt or by accepting the looser behavior.

Test notes

• Tests use n_ctx=6144 instead of the default 8192 because the GDN model plus the embedding model together exceed Metal GPU memory at 8192
• All other non-slow tests pass (31 passed, 1 pre-existing OpenAI API failure unrelated to this change)
• All slow function-calling tests pass (16 passed, 2 skipped for PostgreSQL)