[FEA] `cuvs-bench` should start tracking metrics related to index serialization/write times

As we are starting to evaluate end-to-end performance on serverless architectures like CloudRun w/ GPU, we find that running cuvs bench is a great way to track various build times at different levels of recall on the latency/throughput pareto, however the serialization and writing of the serialized indexes (and even reading of the dataset from the NFS mount) are not being tracked currently and those add to the end-to-end times.

We should start tracking these metrics in cuvs-bench to make it easier to piece together the end-to-end story on architectures like this.

Start by locating cuvs-bench's existing build-time metrics and the stages for index serialization, index writing, and dataset reading. Define how these stages should be timed and reported, then verify that the resulting metrics make end-to-end performance easier to evaluate on the described serverless GPU architecture.