Add PPLX and Jina ColBERT results

Summary

This PR adds complete standard-scope HAKARI-Bench results for two multilingual
late-interaction retrieval models:

Both models have all 563 standard evaluation tasks and all 550 non-overlapping
Overall tasks.

Official settings and validation gate

The official Hugging Face model cards, checkpoint configuration, author blog or
paper, and PyLate usage were reviewed before the full runs.

pplx-embed-v1-late-0.6b uses exact MaxSim over 128-dimensional token
embeddings, [Q] / [D] prefixes, q32/d512, query expansion, and attention
to expansion tokens. It was evaluated in fp32 with SDPA and pinned remote code.
Its NanoBEIR-en mean nDCG@10 was 0.677272, within 0.005149 of the
lightonai/ColBERT-Zero reference used in the official model card.

jina-colbert-v2 uses exact MaxSim, [QueryMarker] / [DocumentMarker],
q32/d300, expansion-token attention, and the documented 128/96/64-dimensional
Matryoshka variants. Its standard Transformers dispatch is eager because the
external XLM-R class does not advertise native flash_attention_2; the pinned
checkpoint's own use_flash_attn: true path uses the installed FlashAttention
implementation. NanoBEIR-en base/96d/64d means were 0.651298, 0.638029, and
0.622335. Across the 12 tasks shared with the official BEIR table, the base
scores correlated at Pearson 0.929994 and Spearman 0.916084.

Runtime

• Hardware: two NVIDIA GeForce RTX 5090 GPUs, one independent worker per
  disjoint task partition.
• PPLX: Python 3.12, Transformers 5.12.1, SentenceTransformers 5.4.1, PyLate
  1.4.0, fp32, SDPA.
• Jina: Python 3.12, Transformers 4.57.6, SentenceTransformers 5.4.1, PyLate
  1.3.4, FlashAttention 2.8.3, bf16, eager standard dispatch plus the
  checkpoint-native FlashAttention path.
• Batch sizes 64 and 128 were used across resumable task partitions. Batch size
  only changes encoding throughput; model weights, token lengths, prefixes,
  dtype, attention, variants, and exact MaxSim scoring were unchanged.

The evaluations were driven by the reviewed local model cards with --all or
equivalent disjoint standard-task partitions. Existing readable result files
were skipped on resume.

Coverage and integrity audit

• 1,126/1,126 compressed result JSON files are present and readable.
• Each model has exactly 563 unique standard task keys, with no missing, extra,
  or duplicate tasks.
• Dataset revisions are resolved and internally consistent.
• All aggregate scores are finite and within [0, 1].
• Every result matches its pinned model revision, dtype, attention setting,
  sequence length, backend, prefix, expansion, and exact-MaxSim metadata.
• PPLX has the expected base result on every task; Jina has base, 96d, and 64d
  on every task.
• A merged-DuckDB audit produced 550/550 Overall tasks for both models. Overall
  base means are 0.570230 for PPLX and 0.551971 for Jina, with Borda ranks
  14 and 25 respectively. Their placement above the established 350M and
  smaller ColBERT baselines is consistent with the NanoBEIR-en validation and
  official BEIR profiles.

Submission checklist

• Only .json.xz result artifacts under the two model directories are
  included.
• No aggregate JSON, caches, DuckDB, logs, Markdown, or scratch files are
  included.
• Model and dataset revisions plus runtime metadata are preserved.
• All 1,126 local files passed decompression, schema, score, coverage, and
  metadata validation.