Bump llama.cpp to b9310 to fix Metal deadlock on macOS 26

[FuJacob]

Summary

Bumps the bundled llama-cpp binary target from b8665 to b9310 (URL + checksum update in Package.swift).

Closes the upstream deadlock that surfaces as Cotabby issue #262: launching Cotabby with Qwen3.5-2B-Q4_K_M.gguf selected hangs the main thread before the menu bar icon appears. Reproduces deterministically on Apple Silicon under macOS 26 (Tahoe).

Root cause

The bundled b8665 Metal backend hits a Tahoe-era IOGPU command-buffer timeout regression. Sampled stack from the affected user:

• Main thread blocked on a pthread mutex waiting for loadModel to finish.
• com.apple.root.default-qos worker spinning forever inside __ggml_metal_rsets_init_block_invoke → usleep → nanosleep → __semwait_signal. That is ggml-metal's residency-set keep-alive heartbeat (introduced in ggml-org/llama.cpp#17766); by design it spins, so the actual hang is the command buffer the loader is waiting on never completing.

Upstream tracked the symptom in ggml-org/llama.cpp#20141 (identical hardware: M4 / macOS 26). The fix was to set AGX_RELAX_CDM_CTXSTORE_TIMEOUT=1 unconditionally inside ggml-metal, which landed around build b8882. b8665 predates that fix. The issue reporter validated b9310 (e2ef8fe42) working on their machine via standalone llama-cli, so that is the target chosen here.

Validation

• swift build (Xcode 26 toolchain): clean.
• swift test: 9 passing, 1 skipped (testEndToEndWithModel, requires COTABBY_TEST_MODEL_PATH). The skipped test is the only one that exercises real model load; reviewers with a local GGUF can run it via COTABBY_TEST_MODEL_PATH=/path/to/model.gguf swift test.
• The C++ wrapper (CotabbyInferenceEngine.cpp) compiles unchanged against b9310, so CotabbyInferenceEngine.h's ABI is unchanged. Downstream Cotabby picks this up on the next package resolve with no source-side edits.

Risk / rollout

• Single-line URL + checksum bump. Reversion is trivial.
• ~675 upstream build numbers (~6 weeks) between b8665 and b9310. No public API removals in that window that the wrapper touches (verified by clean compile of CotabbyInferenceEngine.cpp and full unit test pass).
• Cotabby pins CotabbyInference to branch: main in its project.yml, so merging this PR ships the fix to all Cotabby users on the next build/release without a Cotabby-side change.

Linked

• Cotabby issue: App deadlocks at launch when a Qwen3.5 GGUF is selected (bundled llama.framework hangs in ggml_metal_rsets_init) cotabby#262
• Upstream root-cause issue: Misc. bug: ggml_metal_synchronize error: Mac M4 Pro Tahoe 26.3 Crash due to kIOGPUCommandBufferCallbackErrorInnocentVictim ggml-org/llama.cpp#20141