Change timestamp data type to sfixed64 nanoseconds since Unix epoch

[tigrannajaryan]

The change to int64 is proposed by RFC0059 [1]. In this commit there is a small deviation from
the RFC: here we use sfixed64 instead of int64 to make it consistent with changes already
done for Metrics proto [2]. Experiments proved sfixed64 to be more suitable for timestamps.

[1] https://github.com/open-telemetry/oteps/blob/master/text/0059-otlp-trace-data-format.md
[2] #33

[yurishkuro]

Experiments proved sfixed64 to be more suitable for timestamps.

more suitable than fixed64 too? signed value seems odd for the epoch-style timestamp.

[tigrannajaryan]

Experiments proved sfixed64 to be more suitable for timestamps.

more suitable than fixed64 too? signed value seems odd for the epoch-style timestamp.

fixed64 should be slightly better (double the range in the future), but we already use sfixed64 in metrics. Fro consistency it is best to use the same type for timestamp everywhere. I can update in all places (not sure if we want it in this PR or separate).

[yurishkuro]

If there's no perf benefit of sfixed64, I would go for fixed64, which is unsigned and makes more sense for epoch-style timestamps. Otherwise we'll be answering this question over and over.

[bogdandrutu]

I agree we can change metrics too to use fixed64

[SergeyKanzhelev]

merging this. @tigrannajaryan can you please send separate PR for switch to unsigned everywhere