load_stream loads the whole thing when used intuitively

Surprisingly, load_stream doesn't handle streams but loads all docs in an IO object at once, and only then iterates on the result. This results in much latency as well as huge swaths of memory being taken hostage when used intuitively.

Maybe this could return an Enumerator (that could be to_a) or do a lazy map (that could be forced).

Example "naive" code:

# tested with ruby 2.2.3
require 'yaml'

def rss
  `ps -p #{Process.pid} -o rss`.split("\n").last.to_i
end

unless File.exist? 'test.yml'
  File.open('test.yml', 'wb') do |f|
    1_000_000.times do
      f.write "---\n"
      f.write format("foo: '%016x'\n", Random.rand(2**64))
    end
  end
end

s = YAML.load_stream File.open('test.yml')
puts 'loaded'

s.each.with_index do |_, i|
  puts i if i % 100_000 == 0
end

puts "rss: #{rss} kB"

While load_stream does support a block arg, doing the following is cumbersome and non-discoverable:

s = Enumerator.new do |y|
  YAML.load_stream(File.open('test.yml')) { |doc| y << doc }
end

Review lib/psych.rb around the load_stream implementation and compare its block and non-block behavior using the large YAML example in the issue. Done means intuitive iteration begins processing documents without first loading the entire stream into memory, while the existing block usage remains usable.