oneBehindFetch does not always fetch

This is quite bizarre. For large files, `oneBehindFetch()` does not return any response:

```js
const fetch = require("@ampproject/toolbox-core").oneBehindFetch;

fetch("https://amp.dev/manifest.json")
.then((res) => res.text())
.then((text) => console.log(`length of manifest.json = ${text.length}`));

fetch("https://amp.dev/")
.then((res) => res.text())
.then((text) => console.log(`length of index.html = ${text.length}`));
```

On node 12.16.1, this outputs:

```sh
length of manifest.json = 1136
```

(That is, the length of index.html is not output.) Removing and reordering requests, etc. does not help.

The problem here seems to be that `Response.clone()` (as indirectly used by `oneBehindFetch`) is [essentially broken](https://github.com/node-fetch/node-fetch/issues/151) for rather complicated reasons).

The recommended way of fixing this is to [resolve the cloned responses in parallel](https://github.com/node-fetch/node-fetch/blob/master/README.md#custom-highwatermark), however this is not really suitable for oneBehindFetch's use case, since it endeavours to be a drop-in replacement for `fetch()`.

This is related to #712 and #378. It may also be a factor in #650, since one of the consequences of this is that responses are not cached as expected. (Although that involved a small file, so maybe not.)

Start by locating the oneBehindFetch entry point and reproduce the issue with the Node 12.16.1 example using both amp.dev/manifest.json and amp.dev/. Trace how cloned responses are handled, using the linked node-fetch issue and custom highWaterMark guidance as context. Done means both fetch calls return and their response bodies are readable, without breaking the drop-in fetch behavior.

Written by the indexing model from the issue text.