tiny_psnr: fix range calculation for sample size of 32 bits

For a sample size of 32 bits, the shift would overflow producing
undefined results.  Incidentally, in the only test currently using
32-bit samples, the output matches the reference exactly on most
systems meaning the bad 'max' value is never used.

Signed-off-by: Mans Rullgard <mans@mansr.com>