Memory usage issue with stream-transform

Describe the bug

When using stream-transform for processing large datasets and the parallel option is a value greater than 1, we're seeing high memory usage.

To Reproduce

const fs = require('fs')
const memwatch = require('@airbnb/node-memwatch')
const { pipeline } = require('stream/promises')
const { transform } = require('stream-transform')

let maxUsedHeap = 0

async function main() {
  memwatch.on('stats', (stats) => {
    maxUsedHeap = Math.max(maxUsedHeap, stats.used_heap_size)
  })

  await pipeline(
    function* () {
      let i = -1
      const n = 9999999
      while (++i < n) {
        yield { i }
      }
    },
    transform({ parallel: +process.env.PARALLEL }, (chunk, next) =>
      next(null, chunk.i)
    ),
    fs.createWriteStream('/tmp/output')
  )

  console.log(`${maxUsedHeap / (1000 * 1000)}mb`)
}

main()

// $ PARALLEL=1 node example.js
// 6.009856mb

// $ PARALLEL=2 node example.js
// 320.684144mb

Additional context

• Our theory is that this is backpressure-related. We noticed that this.push was returning false to indicate that the stream should pause reading, yet stream-transform asks for more input over here regardless.
• To support that theory, this seems to resolve the issue, though it isn't a proper solution: https://gist.github.com/477d30dfeb443be9a92ac8a3aedc238f
• Thank you for the CSV project :) its helping us a ton here at snaplet.dev

Start with packages/stream-transform/lib/index.js, especially the input request around line 40 and the this.push call around line 85. Run the supplied large-dataset reproduction with PARALLEL=1 and PARALLEL=2, then inspect the linked gist for comparison. Done means parallel processing respects backpressure without the reported memory growth.