This matches the behavior of net.Socket

Thanks to Stéphan Kochen for the fix and Maurice Fonk for reproducing the
bug.

Closes GH-784.
Closes GH-803.

https://groups.google.com/forum/#!topic/nodejs-dev/V5fB69hFa9o

Closes GH-787.

I have seen a lot of people trying to pass objects to crypto's update
functions, assuming that it would somehow serialize the object before
hashing.

In reality, the object was converted to '[object Object]' which was
then hashed, without any error message showing.

This patch modifies the DecodeBytes function (used exclusively by
crypto at this point) to complain when receiving anything but a
string or buffer.

Overall this should be a less-suprising, more robust behavior.

See: https://groups.google.com/d/topic/nodejs/xJqpp1_s6is/discussion

Data being sent out of order.

Closes GH-774.

http://codereview.chromium.org/6676019/

Closes GH-759.

Closes GH-768.

Change the natives to use this class instead of creating completely new
strings. Reduces memory usage by about 1 MB.

This ensures that the check_function_exists() for fdatasync in the
libeio CMakeLists.txt is not run, since that fails on some versions of
OSX (http://public.kitware.com/Bug/view.php?id=10044).