See PR #306 for details. This is a backport.

Fixed issue whereby element(1, ...) is called on every tuple nested inside a pretty printed record regardless of size.
Modified unit test to test 0 size tuple case.

Adds transparent event stream processing and statistics.
A new 3-tuple trace is introduced as `{Key, Op, Value}`, but
for backwards compatibility `{Key, Val}` implies `=` for `Op`
and `{Key, '*'}` remains as is in the case of wildcards.
A simplified query tree module is generated which reduces
redundant selection conditions to minimize filtering overhead.

This is done via a combination of several things:
* Make the loglevel that triggers a sync configurable
* Make the delayed_write size and intervals configurable
* Make the interval at which external rotation is checked for
  configurable
* Store the timestamp a lager_msg was created inside the lager_msg

To support these changes, several other things had to be modified:
* lager_msg:timestamp now returns a timestamp
* lager_msg:datetime was added to return the {date, time} of a message,
  like lager_msg:timestamp used to
* The configuration syntax for file backends was changed to be of the
  form {lager_file_backend, proplist()} and the old syntax was
  deprecated

Additionally, the defaults for the check_interval was raised from
'always' to 1 second, and the sync_interval was changed from 2 seconds
to one second.

For persistant processes with some immutable metadata (riak vnode and
the vnode ID, for example), implement lager:md/0 and lager:md/1 for
getting/setting such metadata into the process dictionary.

Such metadata is automatically included in any lager message metadata,
so you can just set it in your init() function or whatever and not have
to worry about passing the data around and using it in every lager call.

Originally an idea that Dizzy forwarded to me from Tony Rogvall.
Basicially all the work lager does gathering metadata and stuff is
wrapped in a case statement that fails-fast if the message is definitely
not going to be logged; the severity isn't currently being consumed and
there's no traces installed.

In my simple test, logging 1 million debug messages, when the debug
level is not being consumed, goes from taking 2.99 seconds to 1.21
seconds with this change.

Also, lager pre 1.0 actually had something similar to this, but it was
removed during a refactor and never re-added.

In normal operation, there's no need for log messages to be synchronous.
Its slower and introduces a global serialization point in the
application.

However, when in an overload condition, synchronous logging is good
because it limits each process to at most 1 log message in flight.

So, this change allows the gen_event at the core of lager to switch
modes depending on the size of the gen_event's mailbox. It should
provide better performance in the case of normal load, but it will also
prevent unbounded mailbox growth if an overload occurs.

The threshold at which the switch between async and sync is done is
configured via the 'async_threshold' app env var.

Add add_trace_to_loglevel_config and update_loglevel_config.  These two
handle most of the updates to the loglevel config item, including the
update of the overall logging mask.

This makes minimum_loglevel private.

This doesn't change any behavior -- it just a tidy-up.

Also, adapt the rest of lager to use this bitmask as well.

This also removes the dependency on the syntax_tools and compiler
applications, which interfered with lager's startup/shutdown order.

Helps with pattern matching.

I didn't want the lager_log_message record being used across application
boundaries, this will insulate other applications from any changes to
the message type's internal structure.

Had to add a pseudo-ternary operator to the formatter to support this.
Also allowed lager:log to log metadata.

Changed the messages sent to the backends to include metadata and separated formatting from the backend.  Added documentation, fixed tests, and removed some unused code.