This may seem like a digression at first, but wait for it:
The main purpose of double-blind peer review should not be to prevent the reviewers learning the authors’ identity by all means.
This can be a welcome side effect, but often this is impractical not only due to preprints but also due to previous work, conferences, idiosyncrasies, etc.
If a malignant reviewer wants to break the double blindness, they often can.
Fortunately, many peer reviewers don’t want to do this:
They are completely fine with performing an unbiased review of a paper that is only based on its own merits – as they ideally should.
However, no matter how good our intentions, we are still prone to subconscious or at least implicit biases.
The main goal of double-blind peer review should be to help reviewers to eliminate (or at least reduce) those involuntary biases¹.
By contrast, with single-blind peer review, it is impossible for reviewers to be unbiased, as the names of the authors (and their affiliation) are inescapable.
In light of this, I suggest to submit your (anonymised) paper. In the submission note that you published a preprint already, but that this does not relevantly compromise the double-blindness of the peer review (as it is about allowing the reviewers to be unbiased, not about forcing them). Bonus points if you can point out other ways in which a malignant reviewer could have found your identity despite the preprint. Should the journal not be convinced and actually reject your paper because of this, they should arrive at this conclusion quickly and you waste not too much time.
On the other hand, even if you should somehow manage to unpublish your preprint from the Arxiv, search engines and similar would still connect your paper to your name for a while and thus a malignant reviewer could find out your identity anyway. Therefore, it does not matter for the double blindness at all; no matter what your stance on its purposes is.
¹ … or analogously to prevent authors from flashing their or their affiliation’s reputation at the reviewers without them having a chance to ignore them.