I'm not sure if that's the actual issue but I'm pretty sure __gcd isn't O(1) and it's actually logarithmic which means inside of that triple loop you're calling __gcd millions of times.

I'd say you just need to optimize this but I'm not 100% sure.

Also using vector<vector> dp(n, vector(l, -1e18)) inside of while (t--) might be memory risky

the problem is the gcd probably,__gcd and gcd uses different algorithms, I dont know what algortihm __gcd uses but gcd use Euclidean algorithm which is faster