perf(v2): initialize dataset segment head lazily

[kolesnikovae]

It's been observed that under certain conditions, segment-writer latency might have spikes, even when there are enough CPU resources. The mutex delay profile shows an unusual pattern that fully correlates with the observed latency increase:

I see three potential issues here:

1. An RW lock on a write-intensive path.
2. An expensive defer.
3. A costly allocation under a wide (shard-wide) lock – NewHead is responsible for about 1% of all allocations.

That said, I find it hard to believe these are causing what the profile shows. Even with artificial load tests, I wasn't able to reproduce the problem (the branch version is at least twice as fast in all scenarios). I've seen much more heavily loaded instances that don't show any issues like this. Therefore, I believe it may not fully address the problem.

[alsoba13]

LGTM