Performance: replace linear search with hash-table lookup, so that the O(n*depth) method is now only O(depth)

(but in many cases the caller then calls a O(n) method, so this is O(n) in practice, where n is the number of files in the directory, so no complexity gain in the n=5000 depth==1 case like bug 150885 -- but in practice the remaining O(n) lookup runs much much less code for each iteration than the old code did)