Reduce number of deepcopies in `Pipeline.run` to reduce latency

[sjrl]

Follow up to this issue: https://github.com/deepset-ai/haystack/issues/9011#issuecomment-2854332627

We are probably overusing deepcopying of pipeline and component inputs and outputs especially when setting tags in tracing. It's probably not needed in this instance so it should be explored if it could be dropped.