Reader should handle 0-length reads better

[lilyball]

Reader should be vetted to ensure it handles 0-length reads in a sane fashion. This is inspired by comments on #13049.

A few things need to be ensured:

1. The documentation on Reader discourages returning 0 from read(). It should certainly be allowed, but implementors should be encouraged to avoid returning 0 if possible.
2. The default methods on Reader should all behave sanely if a single call to read() returns 0.
3. Less importantly, the default methods should handle a Reader that returns 0 forever. Practically speaking, this means that if the method employs a loop, it should also employ a counter to count how many 0-length read()s there have been in a row, and if it passes some threshold (TBD, but maybe something like 1000) then it should return an error. There should also be a new IoErrorKind defined for this case (Go uses one called ErrNoProgress, that seems a sensible name).
4. Any clients of Reader in std and the associated libraries should also be checked to make sure they handle 0-length reads.

The simplest way to implement this may be to add another default method read_at_least() that takes a count of how many bytes it should ensure it has read. This method can then deal with counting 0-length reads and setting the proper error, and then everything that reads in a loop can use r.read_at_least(but, 1) so they don't have to deal with it.

[alexcrichton]

I believe this was all fixed in #13127, so closing.