fix: Consistent probability sampler init

[fbogsany]

In the consistent probability sampler, we cache the exponent of the next highest and next lowest negative power-of-two values, and the probability of the highest. E.g. for a probability of 0.1, the next highest negative power-of-two is 2**-3 (0.125) and the next lowest is 2**-4 (0.0625), and the probability of the higher value is the fraction of difference between the 0.0625 and 0.125 that is 0.1 - i.e. 0.1 - 0.0625 / 0.0625 = 0.6.

Unfortunately, we calculated the ceiling incorrectly in the current implementation. Because these are negative power-of-two exponents, the smaller value is the upper bound and the larger value is the lower bound. This corrects that error.

Relevant song: https://www.youtube.com/watch?v=Esdnbq4hxHo

[plantfansam]

👍 Wouldn't mind if we added a test for sampler being initialized with 0.0 (see opentelemetry-go-contrib's testing of the same general stuff). I initially thought this was covered in 62.downto(0) but actually that gets passed to the initializer as 2**-62, which isn't the same thing!

Not essential, but I also think adding a test case where the sampling probability is almost a power of 2 (0.12, 0.48, that kind of thing) can be illuminating to the reader, since you can see a p_ceil_probability that's very apparently skewing towards a given value of p.

[plantfansam]

👍 Wouldn't mind if we added a test for sampler being initialized with 0.0 (see opentelemetry-go-contrib's testing of the same general stuff). I initially thought this was covered in 62.downto(0) but actually that gets passed to the initializer as 2**-62, which isn't the same thing!

Not essential, but I also think adding a test case where the sampling probability is almost a power of 2 (0.12, 0.48, that kind of thing) can be illuminating to the reader, since you can see a p_ceil_probability that's very apparently skewing towards a given value of p.