perf(mcp): disable rate-limiter Redis retries to unblock test suite (#3963)

The MCP rate limiter's @upstash/redis client used the default 5-retry
exponential backoff. In handler tests that point UPSTASH_REDIS_REST_URL at
a fake host, the limiter's EVALSHA falls through to the real network and
burns ~4.3s of backoff per tools/call before degrading open — making
mcp.test.mjs alone ~5.5min and the long pole of `npm run test:data`.

retry:false makes the limiter degrade open immediately on an unreachable
Redis instead of stalling, which is the right behaviour for a best-effort
60/min limiter (never add retry latency to a user request during an Upstash
blip). Enforcement during an outage is unchanged — open either way.

test:data ~7min -> ~1m30s; mcp.test.mjs 5m29s -> 2s; 10,190 tests pass.