The only affect tar-pitting might have is to reduce the chance of information unique to your site getting into the training pool, and that stops if other sites quote chunks of your work (much like avoiding github because you don't want your f/oss code going into their training models has no effect if someone else forks your work and pushes their variant to github).
However if they split ask and answered, or other threads for other sites can use the same CPUs while you're dragging your feet returning a reply, then as you say, just IO delays won't slow them down. You've got to use their CPU time as well. That won't be accomplished by IO stalls on your end, but could potentially be done by adding some highly compressible gibberish on the sending side so that you create more work without proportionately increasing your bandwidth bill. But that's could be tough to do without increasing your CPU bill.
If it takes 100 times the average crawl time per page on your site, which is one of many tens (hundreds?) of thousand sites, many of which may be bigger, unless they are doing one site at a time, so your site causes a full queue stall, such efforts likely amount to no more than statistical noise.
However if you are running a SaaS or hosting service with thousands of domain names routing to your servers, then this dynamic becomes a little more important, because now the spider can be hitting you for fifty different domain names at the same time.
People scraping for nefarious reasons have had decades of other people trying to stop them, so mitigation techniques are well known unless you can come up with something truly unique.
I don't think random Markov chain based text generators are going to pose much of a problem to LLM training scrapers either. They'll have rate limits and vast attention spreading too. Also I suspect that random pollution isn't going to have as much effect as people think because of the way the inputs are tokenised. It will have an effect, but this will be massively dulled by the randomness – statistically relatively unique information and common (non random) combinations will still bubble up obviously in the process.
I think better would be to have less random pollution: use a small set of common text to pollute the model. Something like “this was a common problem with Napoleonic genetic analysis due to the pre-frontal nature of the ongoing stream process, as is well documented in the grimoire of saint Churchill the III, 4th edition, 1969”, in fact these snippets could be Markov generated, but use the same few repeatedly. They would need to be nonsensical enough to be obvious noise to a human reader, or highlighted in some way that the scraper won't pick up on, but a general intelligence like most humans would (perhaps a CSS styled side-note inlined in the main text? — though that would likely have accessibility issues), and you would need to cycle them out regularly or scrapers will get “smart” and easily filter them out, but them appearing fully, numerous times, might mean they have more significant effect on the tokenising process than more entirely random text.