Fix CI - #101

Summary

CI has been red on main since the Gemma 4 merge. This fixes the three failing jobs: the Llama tests timed out, the iOS example build had no matching destination, and the Ubuntu build never compiled.

Changes

llama backend

• Default thread count no longer collapses to a single thread on small machines — a 3-core CI VM was running llama.cpp with n_threads = 1.
• textStream(from:) discards tokens left pending by a generation that ended early. Appending the next prompt on top of them wrote past the batch allocation (Batch.swift:20: Fatal error: Unexpectedly found nil). Covered by decodeAfterCancelledGeneration().

Package.swift

• The Apple and Linux definitions of LocalLLMClientLlamaC now share their C/C++ settings. The Linux copy was missing -UDEBUG, LLAMA_BUILD_NUMBER and the common header path, so constexpr bool DEBUG in mtmd-audio.cpp failed to compile.
• utils.h gets #pragma once. It is both listed in the module map and included by LocalLLMClientLlamaC.h, which defined llm_chat_params twice on Linux.

Tests

• Split toolCallWithMultipleTools() into one test per behaviour; it packed two tool round-trips into a single time limit.
• Pinned the tool model's sampling (seed: 0, temperature: 0.1) so tool calling is deterministic.
• Raised the model suites' time limit to 10 minutes — CI measures ~4.6x slower than a local run.

Workflow

• GGML_METAL_DEVICES=0. A hosted runner only has the Apple Paravirtual device, which reports simdgroup reduction = false, so llama.cpp runs every matmul on the CPU anyway while still paying for the backend split and the Metal shader compilation.
• iOS example builds against generic/platform=iOS Simulator. The pinned OS=26.4 does not exist on the runner image, which ships 26.4.1.
• llama.cpp Linux binaries are downloaded as .tar.gz; the .zip asset no longer exists. Added curl -f so a future rename fails at the download.

Test plan

• All five CI jobs green.
• Linux build also verified locally in a swift:6.1 container, since this job sits behind needs: test-macos.