first benchmark is expensive

I expect identical output for the three benchmarks, but the first one (consistently) needs double time. (Possibly related: #60)

```
import Criterion.Main

main = defaultMain
[ bench "1" $ nf length [1 .. 10^7 :: Int ]
, bench "2" $ nf length [1 .. 10^7 :: Int ]
, bench "3" $ nf length [1 .. 10^7 :: Int ]
]

$ cabal run
benchmarking 1
time 38.37 ms (38.26 ms .. 38.50 ms)
1.000 R² (1.000 R² .. 1.000 R²)
mean 38.67 ms (38.53 ms .. 38.97 ms)
std dev 387.9 μs (201.4 μs .. 535.6 μs)

benchmarking 2
time 20.21 ms (20.13 ms .. 20.31 ms)
1.000 R² (1.000 R² .. 1.000 R²)
mean 20.10 ms (20.03 ms .. 20.17 ms)
std dev 175.3 μs (144.9 μs .. 234.5 μs)

benchmarking 3
time 20.77 ms (20.43 ms .. 21.09 ms)
0.999 R² (0.999 R² .. 1.000 R²)
mean 20.24 ms (20.14 ms .. 20.41 ms)
std dev 306.0 μs (225.8 μs .. 397.1 μs)
```
For criterion-1.6.0.0, with ghc-8.10.7 .. ghc-9.6.1