Document the nuances of dynamic sampling more completely

[cartermp]

Today we don't document some of the subtleties and considerations of dynamic sampling like we could.

@kentquirk mentioned in slack regarding some potentially unexpected behavior about not sampling at goal throughput rates:

Usually because the cardinality of the key fields is too high and/or using too short an evaluation period (and we could explain this better!). Dynamic Sampling always keeps at least one sample of every possible key combination, so if, say, your key field cardinality is 2000, and your cycle time is 30 seconds, and you don't send at least 60K spans/second (30x2000), we can't push sample rates high enough to hit your goals.

There's likely other things that are important for people to know.