Decode has previously been two passes: first chunk the message into a
reversed list of toplevel diameter_avp records, then fold through the
reversed list to build the full result. Various workarounds have made it
a bit more convoluted than it should be however. Rework it completely,
turning the previous 2-pass tail-recursive implementation into a 1-pass
body recursive one.

The relay decode still exists in diameter_codec, as a stripped down
version of the full decode in diameter_gen.

To remove the requirement that dictionary modules be recompiled whenever
the encode/decode implementation changes. The included diameter_gen.hrl
now only contains trivial functions that call info diameter_gen.erl.

To allow list-valued messaged to be encoded in the specified order,
instead of in the dictionary order by first converting the list to a
record. This is not yet exposed in configuration.

By passing additional arguments through it.

Converting with list_to_binary/1 appears to be faster than the
equivalent binary comprehension:

  << (z(F,A)) || {F,A} <- avp_arity(Name) >>

Recursing over the entire list of arities and values is faster than
retrieving them one at a time.

This and subsequent commits are destined for OTP 20.0.