sampling.c and potential divisions by 0 ang log(0) with tablesample and ANALYZE in 9.5

Hi all,
(Petr in CC)

Coverity is pointing out that anl_random_fract and
sampler_random_fract can return 0, causing in some code paths math
errors, aka division by 0 or even log(0) in the case of TABLESAMPLE or
even ANALYZE.

In 9.4, anl_random_fract is careful enough to use random() + 1 to
prevent that, but that's not the case of 9.5 where we begin to use
pg_erand48, that returns a double in range [0.0,1.0).

I think that we should change the returned double to be (0.0,1.0]
instead like in the patch attached (bernouilli and system methods need
a brush-up as well). I haven't updated tsm_system_rows and
tsm_system_time but their regression diffs are attached.

This bug can be triggered when using TABLESAMPLE, now ANALYZE is more
worrying because it could happen during an auto-analyze.
Thoughts?
--
Michael