Generate a Graeco-Latin square

05AB1E, 26 23 22 bytes

-3 bytes thanks to Emigna

-1 byte thanks to Kevin Cruijssen

Lãœ.ΔIôDζ«D€í«ε€нÙgQ}P

Try it online!

R, 164 148 bytes

-many bytes thanks to Giuseppe.

n=scan()
`!`=function(x)sd(colSums(2^x))
m=function()matrix(sample(n,n^2,1),n)
while(T)T=!(l=m())|!(g=m())|!t(l)|!t(g)|1-all(1:n^2%in%(n*l+g-n))
l
g

Try it online!

Dramatically inefficient - I think it's even worse than other brute force approaches. Even for n=3, it will probably time out on TIO. Here is an alternate version (155 bytes) which works for n=3 in about 1 second.

Works by rejection. The function m draws a random matrix of integers between \$1\$ and \$n\$ (without forcing each integer to appear exactly \$n\$ times, or in different columns - this explains the slowness of the code, and is the only thing changed in the faster version). Call this function twice, to get the "latin" and "greek" squares l and g. Then we check:

1. all(1:n^2%in%(n*l+g-n)): are there \$n^2\$ different pairs of values in l \$\times\$ g?
2. are l and g latin squares?

To check that a matrix is a latin square, use the function !. Since point 1. above has been validated, we know that each integer appears \$n\$ times in each of l and g. Compute the column-wise sums of 2^l: these sums are all equal if and only if each integer appears once in each column (and the value of the sum is then \$2^{n+1}-2\$). If this is true for both l and its transpose t(l), then l is a latin square; same for g. The function sd, which computes the standard deviation, is an easy way to check whether all values of a vector are equal. Note that it doesn't work for the trivial cases \$n=0\$ and \$n=1\$ , which is OK according to the OP.

A final note: as often in R code golf, I used the variable T, which is initialized as TRUE, to gain a few bytes. But this means that when I needed the actual value TRUE in the definition of m (parameter replace in sample), I had to use 1 instead of T. Similarly, since I am redefining ! as a function different from negation, I had to use 1-all(...) instead of !all(...).

JavaScript (ES6),  159 147  140 bytes

Returns a flat array of \$n\times n\$ pairs of non-negative integers.

This is a simple brute force search, and therefore very slow.

n=>(g=(m,j=0,X=n*n)=>j<n*n?!X--||m.some(([x,y],i)=>(X==x)+(Y==y)>(j/n^i/n&&j%n!=i%n),g(m,j,X),Y=X/n|0,X%=n)?o:g([...m,[X,Y]],j+1):o=m)(o=[])

Try it online! (with prettified output)

Commented

n => (                      // n = input
  g = (                     // g is the recursive search function taking:
    m,                      //   m[] = flattened matrix
    j = 0,                  //   j   = current position in m[]
    X = n * n               //   X   = counter used to compute the current pair
  ) =>                      //
    j < n * n ?             // if j is less than n²:
      !X-- ||               //   abort right away if X is equal to 0; decrement X
      m.some(([x, y], i) => //   for each pair [x, y] at position i in m[]:
        (X == x) +          //     yield 1 if X is equal to x OR Y is equal to y
        (Y == y)            //     yield 2 if both values are equal
                            //     or yield 0 otherwise
        >                   //     test whether the above result is greater than:
        ( j / n ^ i / n &&  //       - 1 if i and j are neither on the same row
          j % n != i % n    //         nor the same column
        ),                  //       - 0 otherwise
                            //     initialization of some():
        g(m, j, X),         //       do a recursive call with all parameters unchanged
        Y = X / n | 0,      //       start with Y = floor(X / n)
        X %= n              //       and X = X % n
      ) ?                   //   end of some(); if it's falsy (or X was equal to 0):
        o                   //     just return o[]
      :                     //   else:
        g(                  //     do a recursive call:
          [...m, [X, Y]],   //       append [X, Y] to m[]
          j + 1             //       increment j
        )                   //     end of recursive call
    :                       // else:
      o = m                 //   success: update o[] to m[]
)(o = [])                   // initial call to g with m = o = []