Make generate() async to serialize back-to-back turns

[stikves]

Summary

Fixes a crash when reusing an engine for rapid multi-turn conversations:

CoreAIPipelinedEngine.swift:82: Fatal error: Trying to acquire engine when it's already in use

The InferenceEngine.generate() method is now async throws (was throws). The pipelined engine awaits the prior generation's Task before starting a new one, preventing the fatalError on rapid back-to-back calls.

Root cause

PR #64 removed the per-turn engine.reset() call to enable multi-turn KV cache reuse. That reset() also served as the serialization point between consecutive turns (it called drain() internally). Without it, a new generate() call can race with the prior generation's GPU pipeline drain.

Fix

The pipelined engine now cancels and awaits any in-flight generation at the top of generate():

if let priorTask = _generationTask.withLock({ $0 }) {
    _activeToken.withLock { $0?.cancel() }
    await priorTask.value
}

This preserves KV cache state — prefix caching handles reuse automatically across turns.

Test plan

• Existing CancelAPITests pass
• Existing UnifiedGenerationAPITests pass (prefix caching still works)
• New: back-to-back generate() calls do not crash
• New: rapid-fire 10-turn stress test (zero delay between turns)
• New: generate() after cancel() works cleanly
• Manual: multi-turn LanguageModelSession with Qwen3-0.6B pipelined engine

[carinapeng]

Thanks for looking into this! The root cause is helpful tracing :)