Performance Improvement: Don't marshal to proto before the WAL

[joe-elliott]

Describe the solution you'd like
Currently our biggest performance hit is repeatedly marshalling bytes <=> proto. There is technically no reason to unmarshal from bytes to proto before writing a trace to the WAL. We really would only need to do this if the trace were queried, but we do it to every single trace.

Let's remove this marshal/unmarshal for performance. I believe the complexity of this change will come from here:
https://github.com/grafana/tempo/blob/main/modules/ingester/trace.go#L14.

This is the object that we are currently writing and reading from the WAL. To do this we would need to write a different object that contained all of the unmarshalled byte slices.