Improve MergeProfile sample allocations

[metalmatze]

MergeProfiles allocates the worst amount of samples for merging. This might still be possible to improve. While writing this issue, I don't have enough context anymore, but maybe deduplicating the keys upfront and only then allocating could be better? https://github.com/parca-dev/parca/blob/1b2cb213b9f227283cccb7c679fc35acc466476b/pkg/storage/merge.go#L283