Add Sentence Transformers baseline dense results

Summary

This PR adds standard HAKARI-Bench dense --all result files for four
Sentence Transformers models:

Total: 2,204 compressed result files.

Coverage

All four models were evaluated on the standard built-in --all target. Coverage
was audited after rebuilding DuckDB from the submitted result files:

• base rows: 551 tasks per model,
• dense default variants: int8, binary, int8_rescore, and
  binary_rescore,
• each variant has 551 task rows for each model,
• no truncate variants were requested because these models do not document
  Matryoshka/truncate dimensions.

Commands

CUDA_VISIBLE_DEVICES=0 uv run hakari-bench evaluate dense \
  --model sentence-transformers/LaBSE \
  --all \
  --dtype bf16 \
  --attn-implementation sdpa \
  --device cuda:0 \
  --batch-size 128

CUDA_VISIBLE_DEVICES=1 uv run hakari-bench evaluate dense \
  --model sentence-transformers/paraphrase-multilingual-mpnet-base-v2 \
  --all \
  --dtype bf16 \
  --attn-implementation sdpa \
  --device cuda:0 \
  --batch-size 128

CUDA_VISIBLE_DEVICES=0 uv run hakari-bench evaluate dense \
  --model sentence-transformers/all-mpnet-base-v2 \
  --all \
  --dtype bf16 \
  --attn-implementation eager \
  --device cuda:0 \
  --batch-size 128

CUDA_VISIBLE_DEVICES=1 uv run hakari-bench evaluate dense \
  --model sentence-transformers/all-MiniLM-L12-v2 \
  --all \
  --dtype bf16 \
  --attn-implementation sdpa \
  --device cuda:0 \
  --batch-size 128

all-mpnet-base-v2 used --attn-implementation eager because MPNet does not
support the Transformers SDPA attention implementation in this environment.
The other models used SDPA.

Reproducibility

Environment recorded in the result JSON:

• Python 3.12.12
• torch 2.9.0
• transformers 5.3.0
• sentence-transformers 5.4.1
• datasets 4.8.4
• CUDA available, CUDA 12.8
• GPU: NVIDIA GeForce RTX 5090

Validation

• Rebuilt an audit DuckDB from output/hakari-results.
• Confirmed base and dense default variant task counts match for every model.
• Ran uv run --group all pytest tests/test_model_cards.py -q.
• Ran uv run tox.

Checklist

• Result files are committed under hakari-results/{model_dir}/.
• Result files are compressed .json.xz.
• No caches, DuckDB files, HTML reports, or local scratch artifacts are included.
• Result JSON records model revision, dataset revision, runtime configuration, and package versions.
• Core nDCG@10 was generated from the submitted result files.
• Non-default attention choice for all-mpnet-base-v2 is documented.