Cache execution traces (State Compute Cache)

[Stebalien]

Some node operators, especially exchanges and chain explorers, heavily rely on execution traces. Ideally said operators would call StateCompute once, but this isn't always the case.

We can optimize this by caching the result of:

lotus/chain/stmgr/execute.go

Lines 130 to 137 in 97a3b6e

	func (sm *StateManager) ExecutionTrace(ctx context.Context, ts *types.TipSet) (cid.Cid, []*api.InvocResult, error) {
	var invocTrace []*api.InvocResult
	st, err := sm.ExecutionTraceWithMonitor(ctx, ts, &InvocationTracer{trace: &invocTrace})
	if err != nil {
	return cid.Undef, nil, err
	}
	return st, invocTrace, nil
	}

Specifically, we should introduce a simple ARC cache here. It can be small (16-32), but any amount of caching here will likely improve performance. Ideally we'd have a per-tipset lock to avoid duplicating work, but that can be done in a followup if it ends up being expensive.