When calling `#with_request_store` we enable the request store and
ensure it's disabled after the block completes.

However, since the request store is thread local, we can't enable and
disable it twice withing the same block. In regular circumstances,
this doesn't happen, since we only enable the request store in
middleware, and they both run in a different process:

- Using RequestStore::Middleware as a rack middleware for requests
- Using Gitlab::SidekiqMiddleware::RequestStoreMiddleware for
  sidekiq-server.

In specs this could happen when running Sidekiq::Testing.inline!. If
that was used within a spec that had the `:request_store` tag, or
within a feature/request spec that ran the rack middleware.

In this case, the Sidekiq middleware would disable and clear the
request store too soon, causing failures in the specs that counted on
the request store, for example specs counting queries, or checking
the number of Gitaly calls.