Add eval, async_eval, and IO ops (load/save/safetensors/gguf)

[sydneyrenee]

Summary

• eval/async_eval: Force or asynchronously evaluate lazy computation graphs. Accepts single arrays, multiple arrays, or arrays of arrays.
• IO load/save: .npy single-array format roundtrip
• Safetensors: save_safetensors/load with named tensor dicts and optional string metadata
• GGUF: save_gguf/load with named tensor dicts and variant metadata

Build changes

• Swapped IO stub files (no_safetensors.cpp, no_gguf.cpp) for real MLX implementations
• Vendored nlohmann/json v3.11.3 (safetensors JSON parsing dependency)
• Vendored gguf-tools at upstream pinned commit 8fa6eb6 (gguflib.c, fp16.c, bf16.h)

Issues addressed

• Closes Implement mlx.core.eval() #175 (eval)
• Closes Implement mlx.core.async_eval() #176 (async_eval)
• Closes Implement mlx.core.load() #150 (load)
• Closes Implement mlx.core.save() #151 (save)
• Closes Implement mlx.core.save_safetensors() #154 (save_safetensors)
• Closes Implement mlx.core.save_gguf() #155 (save_gguf)

Not included

• Implement mlx.core.savez() #152 (savez) and Implement mlx.core.savez_compressed() #153 (savez_compressed) are Python/NumPy .npz zip format. MLX C++ does not implement them. The same use case (saving multiple named arrays) is covered by save_safetensors.

Test plan

• 10 new tests in batch5-ops.test.ts, all passing
• Full suite: 230 passing, 9 pending (pre-existing)
• eval: single array, multiple arrays, array-of-arrays, async_eval
• npy: save + load roundtrip with shape/value verification
• safetensors: save + load with multiple named tensors, metadata preservation
• gguf: save + load roundtrip with shape verification
• Error handling: non-existent files, invalid paths