Reuse Profiles with stacktraces samples

The biggest memory usage comes from actual samples within profiles. Those are flush once they reach a certain amount default (100_000):

https://github.com/grafana/phlare/blob/arrow-experiment/pkg/phlaredb/profile_store.go#L326-L331

We should be able to reuse then between row group flush but also when creating a new head.

I suspect we should also investigate if we could use a better representation for those profiles and samples like using the parquet.Buffer.
Doing so will also de-amplify serialisation when flushing data.