Currently, the buffer allocated for HttpBody::read_all is of a fixed size. We do check for overflow, but we do this check after writing into the buffer, so we potentially leak returned data somewhere else in memory before returning an error. This PR makes the buffer dynamically-sized, and performs overflow checks earlier.

We also could potentially hit issues with 0-length bodies. Currently, this pattern should never be hit, as we only call this function for ACL lookups and KV store lists, which return non-empty JSON in all paths, so there's no test for this behaviour; it is simply being defensive against future host changes or new uses of read_all.