Proposal
Note - this is a post-1.0 feature.

Issues like JENKINS-41198 and JENKINS-40699 are among the drivers for this - in the Declarative model, parallel doesn't quite fit in smoothly. We need a better answer for this so that more complicated parallel execution of stages is possible within the model.

I'd previously dabbled with a stage execution dependency graph, originally in Plumber, but am easing away from that now. While I love the idea, I can't find a comfortable way to make it work without it being required for every stage, which is a non-option - i.e., I don't think it's a good idea to always require that every single stage has a marker for what stage(s) it can run before or after. So now I'm leaning in the direction of nested stages sections, like this:

pipeline {
  agent any
  
  stages {
    stage('first') {
      steps {
        echo 'first, non-parallel stage'
      }
    }

    stage('top-parallel') {
      stages {
        stage('first-parallel') {
          steps {
            echo 'First of the parallel stages without further nesting'
            sleep 60
          }
        }
        stage('second-parallel') {
          stages {
            stage('first-nested-parallel') {
              steps {
                 echo 'the first of the nested parallel stages'
                 sleep 30
              }
           }
           stage('second-nested-parallel') {
              steps {
                 echo 'the second of the nested parallel stages'
                 sleep 30
              }
           }
        }
     }
  }
}

So in this scenario, stage('first') runs first. When it completes, stage('top-parallel') starts and immediately goes into its nested stages in parallel. stage('first-parallel') starts and goes for 60 seconds, while stage('second-parallel') starts at the same time and descends into its nested stages in parallel as well.

I don't know yet where I'd allow agent and friends, but within a stage, you would need to have one and only one of steps or stages - i.e., a stage either could have steps it executes or it could be a container for parallel stages.

Actually implementing this will need to be done in tandem with Blue Ocean visualization, of course. And this is for now just raw thoughts, but I wanted to get it written down.