ntb_perf: extend with burst/poll/doorbell latency measurement

The ntb_perf test provides a tool for NTB hardware performance
evaluation. For software impact elimination the test uses a simple method
(let's call it 'burst' mode), when the local system sends to the remote
system a data set and counts time interval until hardware completion
report, without the remote side confirming, nor data integrity check.
The measured metric is a 'burst' throughput bandwidth of NTB connection.

The proposed change extends ntb_perf with 3 new metrics. The resulting
test is fully backward compatible.

1. Burst latency. It's a delay between start to send 1 byte to the
remote system and hardware readiness to send another byte. The
measurement performed within bandwidth test procedure. The DMA Engine is
off. Data integrity is not checked. This mode can be disabled by
'perf_latency=N' module parameter.

2. Poll latency. It's a delay between start to send 1 byte to the
remote system and receiving the confirmation. The remote system needs to
be run in server mode beforehand. Then the server polls the input buffer
and on receiving data immediately sends confirmation.

3. Doorbell latency. It's a delay between start to ring an NTB doorbell
and receiving the confirmation. The remote system needs to be run in server
mode beforehand. Then the server waits for a doorbell event and immediately
rings self doorbell to confirm.

Signed-off-by: Alexander Fomichev <fomichev.ru@gmail.com>
Signed-off-by: Jon Mason <jdmason@kudzu.us>