Make response.next lazily computable.

When attempting to debug #4122 I attempted to use the standard method for debugging this, which is to set allow_redirects=False and to manually check the headers. Unfortunately, this no longer works since @kennethreitz shipped 2.15.0, as Requests now always computes the first redirect target in order to write it into Response.next. This makes it very frustrating to debug problems with Requests redirect following because it cannot be disabled, meaning that monkeypatching is required to prevent Requests blowing up.

I propose that we should either make response.next lazily computable, such that it is only calculated when actually requested, or we should remove it.

Expected Result

I expected that if allow_redirects=False was provided Requests would not run any of its redirect handling code.

Actual Result

It did.

Reproduction Steps

import requests
url = 'http://www.yftxt.com/plus/xiazai_yfx.php?open=2&id=12914&uhash=7fadfdb165fadc7362f9dc90'
requests.get(url, allow_redirects=False)

System Information

$ python -m requests.help

{
  "chardet": {
    "version": "3.0.3"
  },
  "cryptography": {
    "version": ""
  },
  "implementation": {
    "name": "CPython",
    "version": "3.6.0"
  },
  "platform": {
    "release": "16.7.0",
    "system": "Darwin"
  },
  "pyOpenSSL": {
    "openssl_version": "",
    "version": null
  },
  "requests": {
    "version": "2.17.3"
  },
  "system_ssl": {
    "version": "100020cf"
  },
  "urllib3": {
    "version": "1.21.1"
  },
  "using_pyopenssl": false
}

Start with the requests.get entry point and the redirect handling that populates Response.next, using the provided allow_redirects=False reproduction. Done means disabling redirects no longer performs redirect handling or computes the first redirect target, with coverage for the resulting Response behavior.