A low effective sample sizes means that the samples the MCMC procedure generated are highly correlated, so they’re not “truly random”. In a way we’re spoiled by having the NUTS sampler; otherwise we’d pretty much always expect to have correlated samples. It could be worse, I’ve sampled models were I run the thing for days and get less than 100 effective samples. It’s not a pathology like divergences are, it just means that you have to be aware that you have less information than you think you do. The rule of thumb I’ve heard is that 1,000 samples is usually enough to do inference and you’re drawing 10k x 0.25 = 2,500, so you should be fine.
As for the future warning, just do what it says, and you can get rid of that az.from_pymc3 as well. I recommend always passing return_inferencedata = True when you sample, because it makes interfacing with Arviz much nicer.