pageserver: relv2 test timeout on arm64 / further optimize read latency

[skyzh]

The test case might take 2m-5m, maybe there's something wrong with the implementation, or the CI machine gets overloaded during some specific time of the day.

[skyzh]

Looking at the trace on my laptop, we spend a lot of time on I/O for the create relation operations:

That's expected: every time we want to create a relation, we will need to issue an I/O to the disk to check if a key (relation) exists or not. And I don't think there's any easy way to optimize it.

Ideas:

• Bloom filter for delta layers so that we don't need to issue I/Os.
• If a user is fully migrated to v2, we can skip the check (Postgres shouldn't give us duplicated keys?)
• Build a cache for the relation dirs.

Our arm CI machines are bound by I/O. I can observe tests take longer time during our peak hour of development (usually morning eastern US time), which makes tests depending on read/write slower. So I'll leave this issue open and look into ways to further optimize it if necessary.