fix(db import): query each PRAGMA individually so checks survive sqlite3 format changes

[ada-powerful]

Fixes #1030.

Problem

turso db import runs sqliteFileIntegrityChecks, which executes one .mode line query and string-matches the resulting text for j = wal / j: wal etc. The dual-separator check covers today's = vs : split, but it depends entirely on whatever line formatter sqlite3 happens to emit. When sqlite3 changes that format again, the check breaks again — and it has already broken once, between 3.49 and 3.50.

Fix

Query each pragma with its own short statement:

sqlite3 -batch -noheader -list <file> "select * from pragma_journal_mode;"

That returns just wal — no key, no separator, no column header — regardless of how .mode line is formatted in any future sqlite3 version. Compare the trimmed value against the expected literal.

Added a readPragma(file, pragma) helper next to runQuickCheck and refactored sqliteFileIntegrityChecks to call it four times. Same error messages, same exit conditions, format-independent.

Tests

$ go test -v ./internal/cmd/...
=== RUN   TestReadPragma
--- PASS: TestReadPragma (0.09s)
    --- PASS: TestReadPragma/journal_mode_returns_wal
    --- PASS: TestReadPragma/page_size_returns_4096
    --- PASS: TestReadPragma/auto_vacuum_returns_0
    --- PASS: TestReadPragma/encoding_returns_UTF-8
    --- PASS: TestReadPragma/nonexistent_file_errors
=== RUN   TestSqliteFileIntegrityChecks
--- PASS: TestSqliteFileIntegrityChecks (0.10s)
    --- PASS: TestSqliteFileIntegrityChecks/valid_WAL_database_passes_all_settings_checks
    --- PASS: TestSqliteFileIntegrityChecks/non-WAL_database_returns_WAL_error
    --- PASS: TestSqliteFileIntegrityChecks/wrong_page_size_returns_page-size_error
=== RUN   TestRunQuickCheck
--- PASS: TestRunQuickCheck (0.12s)
    [all existing subtests pass]
PASS
ok      github.com/tursodatabase/turso-cli/internal/cmd 0.328s

Verified against system sqlite3 3.46.1 in golang:1.25 CI image. Existing TestRunQuickCheck continues to pass.

Disclosure

This PR was prepared by an AI agent (Claude Opus 4.7) under direct human review. Happy to address any feedback.