S6100/Z9100 : Make the get_transceiver_change_event's epoll blocking

[padmanarayana]

- What I did
An optimization was attempted earlier to open the fd and epoll.register during sfputil class instantiation so that only epoll.poll is needed in get_transceiver_change_event(). However, this makes the epoll non-blocking.. (resulting in xcvrd hogging CPU).

- How I did it

1. Create the fd and epoll register during every get_transceiver_change_entry().
2. Do a dummy read so flush outstanding events to prepare for epoll
3. Change the flags to include EPOLLET
4. Check outstanding interrupts that we may have missed (after previous return). Note that since several registers need to be queried, check_interrupts() itself may take 100s of msecs (on S6100, it takes about 350 msecs). In order to ensure that we don't miss interrupts during check_interrupts(), we read the interrupt count before and after check_interrupts(). If there are no fresh interrupts, we block on epoll.poll().

There is still an outstanding race (an interrupt may come after reading the interrupt count but just before epoll.poll() blocks). This time window is about 300 usecs on a S6100. The current workaround is in place until either the fd is made blocking (during class init) or we move to a better mechanism.

- How to verify it

1. Added debug prints and checked that OIR xcvrd blocks and get_transceiver_change_events() is able to construct the dict as expected
2. Checked top output.

- Description for the changelog

S6100/Z9100 : Make the get_transceiver_change_event's epoll blocking

- A picture of a cute animal (not mandatory but encouraged)