Fix: do not sample clock when closing/opening a packet for an event not fitting

Fix: do not sample clock when closing/opening a packet for an event not fitting

This patch adds members to the generated barectf context to save the
current (becomes last) event's timestamp and use it when closing and
opening packets when the event does not fit the current packet instead
of sampling the clock again.

Without this patch, with this scenario:

1. Application calls a tracing function.
2. Tracing function samples the clock.
3. Event does not fit the current packet size: current packet is closed,
   sampling the clock, and new packet is opened, also sampling the
   clock.
4. Event is serialized to the new packet.

The new packet's timestamp is greater than its first event's timestamp,
because this event's timestamp is the result of sampling the clock
before sampling it to close the current packet and open a new one.

Babeltrace 1 accepts and reads the resulting stream, but Babeltrace 2's
muxer is more strict and refuses a message which is in the past.

Signed-off-by: Philippe Proulx <pproulx@efficios.com>
Signed-off-by: Philippe Proulx <eeppeliteloop@gmail.com>