fix(pgtype): JSON DecodeValue returns raw bytes for literal 'null' (#2430)

[luongs3]

Fixes #2430.

Problem

JSONCodec.DecodeValue and JSONBCodec.DecodeValue unmarshal into a Go interface{}, which collapses the JSON document null to a Go nil. That makes a JSON null literal indistinguishable from SQL NULL — both come out of the rows.Values() / Map.Scan(&any) pipeline as nil.

The vanilla database/sql interface does not have this problem, because DecodeDatabaseSQLValue returns the raw response bytes — so QueryRow().Scan(&str) correctly yields "null" for select 'null'::json;.

Repro

rows, _ := conn.Query(ctx, `select 'null'::json;`)
for rows.Next() {
    row, _ := rows.Values()
    fmt.Println(row)
}
// Before: [<nil>]   (looks identical to SQL NULL)
// After:  [[110 117 108 108]]   ([]byte("null"))

Fix

When src is non-nil but c.Unmarshal(src, &dst) yields dst == nil (i.e. the JSON document was literally null, or null with whitespace), return a copy of the raw src bytes instead of the Go nil. SQL NULL (src == nil) continues to return Go nil as before.

Applied to:

• pgtype/json.go — JSONCodec.DecodeValue
• pgtype/jsonb.go — JSONBCodec.DecodeValue (same change, applied after the binary-format version byte is stripped)

What does NOT change

Other JSON values ("hello", 42, [1,2,3], {"k":"v"}) continue to unmarshal into Go values exactly as before. The only behavior change is the JSON-null-but-not-SQL-NULL case, and only to make it distinguishable from SQL NULL.

Tests

pgtype/json_2430_test.go adds TestJSONCodecDecodeValueJSONNullLiteral — 7 table-driven sub-cases exercising the codec directly (no DB required):

• json/sql_null — SQL NULL still returns Go nil
• jsonb/sql_null/text — same for JSONB
• json/json_null_literal — JSON null now returns []byte("null")
• jsonb/json_null_literal/text — JSONB null (text format) returns []byte("null")
• jsonb/json_null_literal/binary — JSONB null (binary format) returns []byte("null") after stripping the v1 header byte
• json/string — "hello" still unmarshals to Go string "hello"
• jsonb/number/text — 42 still unmarshals to Go float64(42)

All pass with the fix; the existing JSON tests that require a live PostgreSQL (TestJSONCodecScanNull, etc.) are unchanged.

[jackc]

As I mentioned in the original issue, I'm not convinced that the current behavior is wrong. But even if it is, having JSON null be special cased to []byte doesn't seem right either.