Investigate removing `AceReplay` from `TraceGenerationContext`

We should look into having the core trace fragment generators generate the `CircuitEvaluation` struct to pass to the `AceChiplet`, instead of having the fast processor include it in `TraceGenerationContext`. The main tradeoffs are size of `TraceGenerationContext` vs speed of `build_trace()` (i.e. what is the performance cost of generating the `AceChiplet` after the core trace, rather than in parallel? Compared to the size increase of `TraceGenerationContext` due to the inclusion of of `CircuitEvaluation`s?)

Context: https://github.com/0xMiden/miden-vm/pull/2188#discussion_r2399162111