Upgrade MLX runtime to 0.32.2

Adopt the new M5 quantized matmul and head-dim-256 attention kernels in the
Darwin dependency floor and bundled Python pins. Refresh the lockfile and keep
packaging/cache contracts source-checked.