Previously, our integration tests were written assuming consistent
reads. To avoid false positives caused by stale reads in low-latency
environments, we enforced consistent reads by setting the consistent
read header. While this reduced noise from false positives, it also
meant we lacked test coverage for replica reads and replication behavior.

To address this gap, we've added a negative test that detects
replication failures by verifying that no stale reads occur after the
expected replication delay.

The new test suite `replica-reads` implements a test that runs 10
trials in parallel. Each trial performs the following steps:
- Set: Writes a random key-value pair to the cache using
`cacheClientWithBalancedReadConcern`.
- Wait: 1 second for reads to propagate
- Get: Reads the value for the key.
- Assert: Confirms that the retrieved value matches the value set
earlier.

Trials are started with slight delays between them to distribute
load (delayBetweenTrials and a small random offset).

Previously we had a separate function in the inner loop of promise
creation. Instead there is a single trial function that is
parameterized by the trial number.

We refactor the test to pull out the trial function and build the
promise in the loop.