Add tokenizer sweep digit variants

[pc0618]

Summary

• Add a Datakit tokenizer sweep DAG for issue Experiment: Measure MoE sensitivity to tokenizer choice #5821 covering GPT-OSS and Llama HF-family tokenizers, both vanilla and place-aligned-digit variants.
• Train 262k vocabularies on a deterministic 50B-token-equivalent sample, then derive 128k, 32k, and 8k BPE tokenizers by rank truncation.
• Add the bounded issue Number tokenization #4915 numeric pre-tokenizer: isolate numeric runs, split right-to-left into place-aligned 3-digit groups, and cap regex digit runs at 510 chars to avoid catastrophic backtracking.
• Add retokenization/metadata plumbing for train and holdout caches.

Tests

• python3 -m py_compile experiments/datakit_testbed/tokenizer_sweep_20260526.py tests/datakit_testbed/test_tokenizer_sweep_20260526.py
• git diff --cached --check before commit
• uv run pytest tests/datakit_testbed/test_tokenizer_sweep_20260526.py is currently blocked locally during repo conftest import because the unsynced environment cannot import workspace package fray (ModuleNotFoundError: No module named 'fray').