fix: inject fallback model name into CompletionRequest

[voidborne-d]

Problem

When FallbackDriver retries a request with a different provider, request.model still carries the primary model name. This causes the fallback provider to receive an invalid model ID (e.g., sending claude-3-opus to an OpenAI endpoint).

Fixes #200

Root Cause

FallbackDriver cloned the request as-is for every driver in the chain. The model field was never updated to match the fallback entry's configured model.

Solution

• Added FallbackEntry struct that pairs each driver with an optional model override
• New FallbackDriver::with_models() constructor accepts Vec<(Arc<dyn LlmDriver>, String)>
• Both complete() and stream() now inject the correct model before forwarding
• Updated both kernel boot path and resolve_driver() to use with_models()
• Existing FallbackDriver::new() is preserved for backward compatibility
• Added test test_with_models_overrides_model_on_fallback verifying the override

Changes

• crates/openfang-runtime/src/drivers/fallback.rs — new FallbackEntry, with_models(), model injection in complete()/stream()
• crates/openfang-kernel/src/kernel.rs — both fallback construction sites now pass model names