cache column metadata for prepared and cached statements

[mattn]

Column names and declared types are invariant for the lifetime of a compiled statement, but Columns() / declTypes() called C.sqlite3_column_name and C.sqlite3_column_decltype once per column on every query. On hot QueryRow paths that reuse an explicit db.Prepare(...) or a _stmt_cache_size cached statement (#1387), this is a fixed set of cgo crossings paid on every execution.

This caches the names and decltypes on the SQLiteStmt the first time they are materialized (lazily, under the existing statement mutex) and reuses them afterwards. Caching is gated by cacheMetadata(): explicit prepared statements always cache, Query-created ephemeral statements cache only when they live in the connection statement cache, and one-shot statements keep the previous per-call behavior. No new allocations on the steady-state hit path, no behavior change for non-cached statements.

Read-only point lookup (SELECT id, name, hp, attack FROM monster WHERE id = ?, 1000-row DB, cache=shared&mode=ro&_journal_mode=WAL, _stmt_cache_size=32), linux/amd64 Ryzen 7 7735HS WSL2, -benchtime=400ms -count=3 median. C1/C8 = SetMaxOpenConns; P10 ≈ 160 goroutines.

Effect of this change (cached statement path):

Condition	master	this PR
C1 Seq	7.9 µs	6.6 µs	-16%
C1 P10	23.4 µs	18.6 µs	-20%
C8 P10	12.9 µs	12.3 µs	-5%

The win is largest at low / single-connection concurrency, where every cgo crossing sits on the critical path, and shrinks as connection parallelism hides the saved crossings.

For context, best config of each driver on the same workload — mattn = _stmt_cache_size=32 + this change + SetMaxOpenConns(8), modernc = modernc.org/sqlite with unlimited connections (its fastest setting here). P1/P10/P100 ≈ 16 / 160 / 1600 goroutines:

Condition	mattn	modernc
Seq	6.3 µs	13.8 µs	mattn 2.2x
P1	12.1 µs	12.5 µs	~tie
P10	11.5 µs	17.1 µs	mattn 1.5x
P100	13.2 µs	~115 µs	mattn

With the statement cache, this metadata cache, and a connection pool sized below the cgo thread-pinning ceiling, mattn is competitive-to-faster across the board on this workload.