RedshiftQuery: negative offset on custom granularity produces invalid SQL (-- line)

**Describe the bug**

When a custom granularity is defined with a **negative** `offset` (e.g., `offset: -1 day` as documented in Cube.dev's own [sunday_week recipe](https://cube.dev/docs/guides/recipes/data-modeling/custom-granularity)), the Redshift dialect emits invalid SQL. The inner `DATEADD` renders `--1`; two consecutive minus signs, which Redshift parses as a **SQL line-comment start**. The rest of the line is treated as a comment, so the query fails with:

```
syntax error at or near "FROM" in context "(day, --1, ..."
```

**Root cause** `subtractInterval` in `RedshiftQuery.ts` uses literal string concatenation to negate the interval value, with no parenthesization:

```typescript
result = `DATEADD(${datePart}, -${intervalValue}, ${result})`;
```

When `intervalValue = -1`, the template concatenates `-` + `-1` producing `--1`.

Source: [`packages/cubejs-schema-compiler/src/adapter/RedshiftQuery.ts`](https://github.com/cube-js/cube/blob/v1.7.4/packages/cubejs-schema-compiler/src/adapter/RedshiftQuery.ts) at v1.7.4, method `subtractInterval`.

**To Reproduce**

1. Define a cube with a time dimension and a custom granularity with a negative offset:

```yaml
cubes:
- name: orders
sql: >
SELECT 1 AS id, DATE '2026-01-07' AS event_ts, 100 AS amount
UNION ALL SELECT 2, DATE '2026-01-14', 200
UNION ALL SELECT 3, DATE '2026-01-21', 300
dimensions:
- name: event_ts
type: time
sql: "{CUBE}.event_ts"
granularities:
- name: sunday_week
interval: 1 week
offset: -1 day # ← the buggy case
measures:
- name: total
type: sum
sql: "{CUBE}.amount"
```

2. Configure Cube against Redshift as the data source.
3. Query the cube via the Data API:

```json
{
"measures": ["orders.total"],
"timeDimensions": [{
"dimension": "orders.event_ts",
"granularity": "sunday_week",
"dateRange": ["2026-01-01", "2026-02-01"]
}]
}
```

4. `GET /cubejs-api/v1/sql` shows the emitted SQL contains the bug:

```sql
SELECT DATEADD(day, -1, date_trunc('week', DATEADD(day, --1, ...)))
^^^^^ SQL line comment starts here
```

5. `GET /cubejs-api/v1/load` fails: `Error: syntax error at or near "FROM"`

**Expected behavior**

The inner `DATEADD` should render as valid SQL, e.g. `DATEADD(day, 1, ...)` ; the negation computed as `-(intervalValue) = -(-1) = 1`. Full expected SQL:

```sql
SELECT DATEADD(day, -1, date_trunc('week', DATEADD(day, 1, ...)))
```

**Minimally reproducible Cube Schema**

See "To Reproduce" above. Same bug reproduces regardless of yaml vs JS schema format, the defect is in the SQL emitter, not the schema parser.

**Version:**
1.7.4 through 1.7.16 (latest as of 2026-07-31). Verified byte-identical at [v1.7.4](https://github.com/cube-js/cube/blob/v1.7.4/packages/cubejs-schema-compiler/src/adapter/RedshiftQuery.ts) and on `master`.

**Additional context**

**Suggested fix** two clean options in `RedshiftQuery.ts`:

```typescript
// Option A: parenthesize
result = `DATEADD(${datePart}, -(${intervalValue}), ${result})`;

// Option B: compute negation in JS (cleaner and never emits leading `--`)
result = `DATEADD(${datePart}, ${-intervalValue}, ${result})`;
```

The same fix pattern likely applies to `addInterval` if any other code path can pass a negative value that becomes doubly-negated.

**Related** the DuckDB dialect had the same bug class, fixed in #11272 (v1.7.3, 2026-07-16). Redshift's `subtractInterval` was not touched by that PR. A wider audit of the other dialect overrides of `subtractInterval` / `addInterval` (Postgres uses `+/- interval '...'` and is safe; MSSQL, BigQuery, Snowflake, Athena, Databricks etc. should be checked) is likely worthwhile.

**Workaround** use a positive offset that's mathematically equivalent modulo the interval length. For a 1-week interval, `offset: 6 days` ≡ `offset: -1 day` (both anchor buckets to Sunday). This avoids the `--` string entirely. Verified empirically on Cube Cloud 1.7.4: 10 Sunday-anchored buckets returned, no SQL syntax error.

Start in packages/cubejs-schema-compiler/src/adapter/RedshiftQuery.ts at subtractInterval, then reproduce the issue with the provided custom-granularity schema and the Data API SQL endpoint. Done means a negative offset emits valid Redshift DATEADD SQL without a leading -- and the example load request no longer fails; also inspect addInterval as suggested for the same case.

Written by the indexing model from the issue text.