Ann-bench: fix unsafe lazy blobs

[achirkin]

The ann-bench dataset uses lazy-loading blobs to move data between storage and host and device memory.
The data may be moved between memory spaces at the moment some properties/pointers are requested.

In the search throughput mode, this sometimes causes a problem: two concurrent benchmark threads access the same property and concurrently modify the state of the blobs, which leads to various segfaults.

The fix is to guard the critical sections with a mutex lock. There shouldn't be any impact on benchmark QPS results.
Only one method, dataset->dim() is accessed within the benchmark loop. To avoid locking the mutex in this method, this PR changes the way dim() is evaluated; it's cached in dim_ variable while still maintaining the behavior of loading it either from the query set or the base set depending on what is available/accessed first.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 84.00%. Comparing base (b801218) to head (2b43309).
Report is 4 commits behind head on branch-25.06.

Additional details and impacted files

@@              Coverage Diff              @@
##           branch-25.06     #828   +/-   ##
=============================================
  Coverage         84.00%   84.00%           
=============================================
  Files                18       18           
  Lines               125      125           
=============================================
  Hits                105      105           
  Misses               20       20           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tarang-jain]

Thanks for the update. It is a smart idea that we can put locks on the dataset layer (dataset.hpp) and still ensure thread safety in helpers in the blob layer (blob.hpp). IIRC currently the blob header is not used anywhere directly apart from dataset.hpp, so I do not have any concerns with this change.

[tfeher]

Hi Artem, thank you for this fix! Overall it is fine, but there is one issue.

[tfeher]

Thanks Artem for the updates. This works, but do we need all the added complexity?

When I asked to cache dim, I was only thinking of to store it in a variable outside the benchmark loop (like it is done for query_set).

[achirkin]

I actually thought this is a little bit of a simplification :) two variables base_set_accessed_, query_set_accessed_ are replaced with a single dim_, and a helper function in place of plain assignments clearly states what is it doing for the program logic.

[tfeher]

I actually thought this is a little bit of a simplification

Ok, I accept your point above. Still I believe the helper function is doing more than what it should.

[tfeher]

Thanks Artem for the updates, LGTM!

[tfeher]

/merge