Fix MIRACL eval leakage via dev split and add full English MTEB evaluation (bf16 + FlashAttention-2)

[iamtatsuki05]

WHY

The alignment/uniformity and geometry analyses loaded miracl/miracl with a hard-coded split='train', which overlaps with the FT training data (ruri-v3-dataset-ft's miracl subset is built from the same train queries). In addition, the paper's English MTEB numbers covered only one task per family, so we re-run the full MTEB(eng, v2) task families under the same precision/attention setup used for training.

WHAT

MIRACL leak fix (alignment / uniformity)

• Expose miracl_split (and positive-pair split) in the eval config and default all evaluation positive-pair splits to dev
• Pass prepare_dataset arguments by keyword to avoid positional drift

Isotropic eval harness hardening

• Add dtype / attn_implementation fields to CLIConfig (bfloat16 with flash_attention_2 or sdpa)
• Load models with trust_remote_code so bidirectional remote-code models do not silently fall back to causal attention, and validate the reported attention backend / dtype
• Enforce the 8192-token contract when the model supports it; measure short-context models at their native lengths with a warning
• Add English base-model (pre-training only) isotropic configs

Full English MTEB evaluation

• Add per-model full MTEB(eng, v2) configs (37 tasks across the 5 reported task families) and 120 chunked configs for parallel scheduling
• Load MTEB eval models in bfloat16 (FlashAttention-2 rejects fp32)
• Sanitize empty texts at the tokenize level (FiQA contains empty documents, which crash tokenizers that do not force special tokens; MTEB passes numpy arrays, so the sanitization hooks tokenize rather than encode)

Misc

• Apply ruff format to the isotropic eval script and sentence-model config