I am not particularly surprised. The number in the table is also the number of times the loop will iterate in the simple implementation and literally half of the table entries are zero. Of the remaining half, half are one. Of that remaining half, half are two, and so forth. The cost of a memory read on modern hardware seems to be high enough that the lookup table does not give enough of an advantage. Perhaps the bit-twiddling solution roboticus offered will work better for you?
Or is your testing script swamping the function under test with Perl-side type conversion and loop overhead? I suggest moving the iteration into C instead of walking a Perl array, since I suspect that both of the solutions in your test probably have running times dominated by the Perl for iterating over the array and the XS magic that converts an SV into a UV. Moving that iteration into C is probably necessary in any case: if you will be calling this predicate from Perl code, you need to know the overhead for that call itself to determine if optimizing uv_fits_double will actually provide any overall performance gain.
Also note that uv_fits_double3 as presented here enters an infinite loop if given zero as an argument.