I copied your solution to custom invocation, and tried to simplify as much as possible without changing the verdict. Eventually, I realised: the culprit is

fr1(i,1,n-1){
    for(auto j:bol[i]) coef[i]+=1ll*j.se*bol[i+1][j.fi];
}

Here you essentially create quadratic memory by creating entries in the bol[i] for all numbers, that are present in any row from 1 to i-th — remember that even access to a std::map creates an entry. So, if n = 1e5, m = 1, and all numbers are distinct, than you create approximately n*n/2 entries in the maps.