lfm2 : make LFM2.5-VL-3B work with a DSpark draft model

[tugot17]

This makes LFM2.5-VL-3B work with a DSpark draft model.

The lfm2 graph lost its t_layer_inp capture in the models/ split (still there at 38a5b42, and qwen3.cpp kept its copy). Without it the server asserts on the first request whenever a drafter is attached:

llama-graph.cpp:1352: GGML_ASSERT(t_layer_inp[il] != nullptr && "layer input tensor is null")

One line to put it back.

Testing:

hf download LiquidAI/LFM2.5-VL-3B-GGUF LFM2.5-VL-3B-F16.gguf mmproj-LFM2.5-VL-3B-F16.gguf --local-dir models
hf download tugot17/mango-gguf mango-draft-F16.gguf --local-dir models

./build/bin/llama-server -m models/LFM2.5-VL-3B-F16.gguf \
  --mmproj models/mmproj-LFM2.5-VL-3B-F16.gguf \
  -md models/mango-draft-F16.gguf \
  --spec-type draft-dspark --spec-draft-n-max 8 --spec-draft-n-min 0 \
  -fa on -ngl 99 -c 8192

Target and mmproj are public, the draft sidecar is gated on my HF, ping me for access.

Accept length on a 20-sample MMSpec probe: 3.86 (at 38a5b42), our internal fork gets ~4.1 on the same samples. Greedy output is identical with and without the drafter.

Things I hit on the way that this PR does not fix:

• accept length dropped ~30% for every DSpark draft between 38a5b42 and current master (3.86 -> 2.85, same probe). needs a bisect, my guess is spec: support speculators-format checkpoints for DSpark ggml-org/llama.cpp#26275
• the Lfm2DSparkDraftModel converter path is gone from master, I converted the sidecar with an older tree. btw the old converter applied the q/k rope permute twice for rope_is_neox_style=false checkpoints, which is why all our earlier VL sidecars measured ~1.7. it has to be applied once
• convert_hf_to_gguf --target-model-dir doesn't accept a VL target dir, Lfm2VlForConditionalGeneration is only in the mmproj map