Add oi_parser parse provider and pipeline

[thiagosantos1]

This resubmits #77, which was closed because oi-parser API keys required manual admin approval (not self-serve). That's now fixed — details below. Opening a new PR since closed PRs can't be reopened.

---

Adds an oi_parser parse provider + pipeline for the hosted oi-parser document-parsing API (POST /v1/extract, X-API-Key auth, returns markdown + layout). Also includes a layout cross-eval adapter + label mapper so the pipeline can be scored on the layout-detection dataset.

Getting an API key: sign up at https://oi-parser.ai/ — keys are now issued instantly and active on signup, with no admin review or "Pending approval" step. One active key per user, valid 14 days, and a new key can be self-issued anytime — also with no approval. Set it via OI_PARSER_API_KEY (optionally override the endpoint with OI_PARSER_BASE_URL).

Recommended run flags:

• --max_concurrent 20 — the per-API-key concurrency cap on the default key; exceeding it only wastes retries. Raise it if your key is provisioned with a higher limit.
• --per_file_timeout 900 — the service allows up to ~900s per document; the default 600s can cut off the slowest documents.

Example:

parse-bench inference run oi_parser --input_dir <dir> --max_concurrent 20 --per_file_timeout 900
parse-bench evaluation run <output>

[thiagosantos1]

@boyang-zhang1 Please let me know if there's anything missing from our side to move the PR forward