deps: bump llama.cpp correctness backports

[davide221]

Follow-up hub bump for Luce-Org/llama.cpp-dflash-ggml#21.

This updates server/deps/llama.cpp from the RDNA3.5 MMQ override merge to the
current luce-dflash head:

• old: 6fd3d84e2168476d5e199a2fa1221d82ba883c21
• new: 30c9d7dac8fee3c6c4cb1dddf7b97c0355cab00f

Included llama.cpp changes:

• CUDA/HIP flash-attention MMA mask offset overflow fix.
• CUDA GGML_OP_REPEAT support check restricted to implemented F32/F16 paths,
  avoiding runtime asserts for unsupported types.

This PR intentionally only changes the submodule pointer.

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}