fix(metal): add turbo2/3/4 types to FLASH_ATTN_EXT and CPY support ch…

[tomwolfe]

Fix: Enable GPU offloading for TurboQuant KV cache types in Metal backend

• Issue: GGML_OP_FLASH_ATTN_EXT and GGML_OP_CPY were falling back to the CPU for TurboQuant KV cache types due to missing support in the Metal backend's supports_op check.
• Solution: Updated ggml/src/ggml-metal/ggml-metal-device.m to explicitly support the following types for these operations:
  • GGML_TYPE_TURBO2_0
  • GGML_TYPE_TURBO3_0
  • GGML_TYPE_TURBO4_0
• Result: Operations are now correctly offloaded to the GPU.
• Verification: The project builds successfully with these changes.