If the bad guy has intercepted the DNS, they can just provide the md5 of the bad script, no?
That's enough extra work, and unreliable enough, that the attacker might not bother. Why work so hard to sabotage the user who checks md5sums when you can just wait for a user that doesn't? Just because thieves can carry lockpicks doesn't mean that you shouldn't bother locking your car: Protection against lazy, opportunistic thieves is still better than nothing.
The other advantage of the MD5 plan is that you can download the MD5 from a different site than the script, at a different time and over a different internet connection (or, perhaps, over https). A specific, important version of that use case is: If you're installing the script over and over again in an automated fashion, you can download its MD5 in advance, cache it, and then check it against every future download of the script to verify that the script hasn't changed. When the script gets updated and the MD5 legitimately changes, you audit the diff and then update your copy of the MD5 for the future.
Meanwhile, using curl-over-HTTPS seems like it couldn't hurt, but better make sure 'curl' is really checking the cert and aborting on cert mismatch, because tools can be very sloppy about this. Also, you're still trusting the third party site, and once their site gets hacked it's game over… unless you have another canonical source for the MD5 sum.
One ultimately realizes why real packaging systems have signed packages, with private keys assigned to developers.
Maybe the solution here, assuming you trust the third party, is for them to get a signed SSL cert and provide `curl https://get.whatever.com|bash`