Use a per-request DuckDB connection

[gordonmurray]

Closes #9. Refs #10.

Problem

DuckDBManager kept one shared connection at module scope, and _apply_s3_config closed it and opened a new one on every request. Two concurrent requests could collide: request B's reconfigure would close the connection request A was mid-query on, surfacing as spurious errors or (worse) cross-contaminated session state.

Fix

Per-request connection lifecycle. No shared state.

• _duckdb_connection(config) — a context manager that opens a fresh in-memory DuckDB, installs+loads extensions, applies the caller's S3/Iceberg config, and closes on exit.
• _apply_s3_config, _attach_iceberg_catalog, _validate_iceberg_table, _convert_to_iceberg_query are now plain functions that take a conn argument. No self.connection. No hidden mutation.
• Route handlers (test_connection, execute_query) delegate to run_connection_test / run_query impls — renaming avoids the route/impl name clash.
• DuckDBManager class and the db_manager global are gone.
• Startup lifespan now just pre-installs extensions once so the first real request doesn't pay the download cost. DuckDB caches the binaries on disk, so each per-request INSTALL/LOAD is ms-level after that.
• Dropped PRAGMA enable_profiling=json and PRAGMA profiling_output='query_profile.json' — the JSON file was never read by any code, and writing a shared filename from many connections would reintroduce the exact race we're fixing.

Test suite

Since the refactor made the validators and converters module-level (testable without a live DB), a second commit adds a pytest unit suite and wires it into CI:

• backend/tests/test_sql_validation.py — SELECT/WITH/UNION accepted with LIMIT injection, preserves existing LIMIT, rejects INSERT/DROP/DELETE/UPDATE/CREATE/ATTACH and multi-statement input
• backend/tests/test_query_conversion.py — read_parquet() → iceberg_scan() and REST catalog rewriting
• backend/tests/test_connection_config.py — Pydantic validators: SQL-injection attempts in tablePath, non-identifier namespaces, bad regions, control chars in credentials, /metadata stripping
• backend/tests/test_sql_string_literal.py — quote doubling, control-char rejection, tab allowance
• conftest.py + pytest.ini for clean discovery without PYTHONPATH
• New backend-tests CI job (setup-python 3.12, pip cache, pytest)

35 tests, ~0.5s runtime, no DuckDB engine / network / docker needed.

Verified locally

• ruff check backend/ scripts/ clean.
• pytest 35/35 pass in 0.48s.
• End-to-end via docker compose up --build: demo query returns 37,537 rows.
• Concurrency test — 5 parallel requests with different queries and row limits each return their own correct row count:

A: rows=5    (LIMIT 5)
B: rows=1    (COUNT(*))
C: rows=23   (LIMIT 23)
D: rows=13   (GROUP BY decade)
E: rows=7    (LIMIT 7)

Before this PR the same test would intermittently fail or return wrong counts as connections got swapped mid-query.

• DROP TABLE foo still rejected with 400 Only SELECT queries are allowed (got DROP) — sqlglot validator unchanged.