This is one of the self-contained threads to unravel for cockroachdb#18784.

This is also required for cockroachdb#18784.

See cockroachdb#18779.

This doesn't carry out all the work but does all the basic required refactoring (mod cleanups).

I don't love that we're passing `ReplicaEvalContext` through an interface. That should be avoidable
and doing so should be instructive (I suspect we'll trim some fat in the process):

Many methods roughly take the following form:

```go
// GetMVCCStats returns the Replica's MVCCStats.
func (rec ReplicaEvalContext) GetMVCCStats() enginepb.MVCCStats {
    r := rec.repl
    r.mu.RLock()
	defer r.mu.RUnlock()
	return *r.mu.state.Stats, nil
}
```

and to define this method inside of the `batcheval` package, we'd have instead

```go
type ReplicaEvalContext {
    // ...
    state struct {
        *sync.Mutex
        *storagebase.ReplicaState
    }
}
```

though having to copy the mutex pointer smells fishy. Similar strategies can be used for most of the
other methods. The downside is that `ReplicaEvalContext` will wind up being a struct with lots of
pointers (i.e., perhaps a little large to sit in the hot path, though much smaller than
BatchRequest!). We could accommodate that by arranging some of these pointers that never change (for
example the `Engine` or the `AbortCache`) into a long-lived struct on `*Replica` instead.