deps: bump llama.cpp - RDNA3/RDNA4 MMQ tile override (+6-8% decode)

[geometric]

Summary

Bumps the server/deps/llama.cpp submodule (luce-dflash) to pick up the
RDNA3/RDNA4 MMQ tile override for DFlash's small spec-decode verify batches.

The change uses a 48x64/4-warp tile in place of the stock 128x128/8
tile and keeps output bit-identical in the measured decode workloads.

Impact

Scenario	before	after	gain
gfx1201 (R9700), Qwen3.6-27B Q4_K_M, --ddtree-budget=22, n_gen=256	54.65 tok/s	59.37 tok/s	+8.3%
gfx1100 (RX 7900 XTX), Qwen3.6-27B Q4_K_M, --ddtree-budget=22, n_gen=256	56.78 tok/s	60.18 tok/s	+6.0%

Implementation

• Bumps server/deps/llama.cpp to a luce-dflash commit containing the smaller RDNA MMQ tile.
• Documents the HIP performance context in server/docs/HIP_PERF_PLAN.md.
• Leaves gfx1151/RDNA3.5 on the default path because it was not benchmarked.

Runtime Flags / Configuration

No new runtime flags or configuration switches.

Validation

Check	Result
10-prompt HE mean, Qwen3.6-27B Q4_K_M, --ddtree-budget=22, n_gen=256	Output bit-identical; gfx1201 improves 54.65 -> 59.37 tok/s, gfx1100 improves 56.78 -> 60.18 tok/s.

Notes

Depends on Luce-Org/lucebox-ggml#18 for the underlying kernel
change. This branch points the submodule at that PR's head commit; before
merge, repoint the SHA to the luce-dflash merge commit of Luce-Org#18.

This is marked draft for testing the GitHub App manual PR flow.