fix(curation): remove redundant wrapper parse in reject_non_single_select

[amansingh-121]

Closes #481.

Removes the redundant third parsing step in reject_non_single_select (src/hflow/curation.py). The function already rejects multi-statement inputs and PRAGMA/DESCRIBE/SHOW/SUMMARIZE inputs via DuckDB's extract_statements count/type check and the leading-keyword check, respectively. The wrapper parse SELECT * FROM (<sql>) never catches a case that isn't already refused, and it adds a per-query DuckDB parse for no benefit.

The wrapper parse was originally added to catch subquery-forbidden leading keywords, but those are already covered by the _leading_keyword gate, and any SQL that parses standalone as a single SELECT also parses inside the wrapper (or is already rejected by the count/type check). Measurements in the issue confirmed the block is dead: removing it leaves all 9 reject_non_single_select tests green, and a broad shape sweep found no input where the wrapped parse fails while the first two checks pass.

Changes

• Remove the wrapped_input construction and the extract_statements(wrapped_input) try/except.
• Update the docstring and inline comment to remove references to the wrapper parse.

Tests

• All test_reject_non_single_select* tests pass (9/9).
• ruff check, ruff format --check, and ty check all pass.

Notes

The change is purely a dead-code removal — no new behaviour is introduced, and no existing refusal messages change. The per-query parse cost drops by one DuckDB parse for every curation gate check.

[kstonekuan]

Please stop taking good first issue tickets. You were asked on #518 and this is another one.

Not reviewing this.