reference: Heun with 6000 steps, 11999 denoiser calls

median |error| of the final sample against the reference, and share of samples that landed in the wrong mode
steps       euler NFE  wrong        heun NFE  wrong    dpmpp_2m NFE  wrong  
5        0.466613 5    0.0512   1.139991 9    0.0132   0.398274 5    0.0092 
10       0.151898 10   0.0328   0.150810 19   0.0003   0.215795 10   0.0053 
20       0.074535 20   0.0135   0.024285 39   0.0010   0.033782 20   0.0000 
40       0.036909 40   0.0055   0.004994 79   0.0005   0.006319 40   0.0000 
80       0.018398 80   0.0022   0.001164 159  0.0003   0.001394 80   0.0000 
160      0.009174 160  0.0003   0.000281 319  0.0000   0.000330 160  0.0000 
320      0.004581 320  0.0000   0.000069 639  0.0000   0.000080 320  0.0000 
640      0.002290 640  0.0000   0.000017 1279 0.0000   0.000020 640  0.0000 

measured order: slope of log(median error) against log(steps)
sampler    10-40   40-160  160-640
euler      1.02    1.00    1.00
heun       2.46    2.07    2.03
dpmpp_2m   2.55    2.13    2.04

same budget of about 20 denoiser calls, distance of the whole sample set from the data
euler      20 steps   20 calls  W1 = 0.1205
heun       10 steps   19 calls  W1 = 0.1487
dpmpp_2m   20 steps   20 calls  W1 = 0.0384
floor, two real samples of this size: W1 = 0.0189
