The tagger is for English and I will use the universal dependency tag set. I am looking for a corpus that best represents nowadays English.
Whichever corpus is reasonably large and correctly annotated and has similar content to what the tagger will actually be used on. (Maybe you can be more specific about what you need.)
Realistically you may want to train on or find a pretrained model trained on a more general corpus that is larger, and the fine-tune on a presumably smaller more targetted corpus.
Another way to stretch a small targetted corpus is to use representations from general models trained without supervision.
I assume spaCy has instructions for using it with those approaches, and that it's still more developer friendly than anything from Stanford or Google.
You will probably not be best served by seeking a tagger that can handle "all of English". It is better to consider the kinds of English that your tagger will be encountering (e.g., literature, Wikipedia, newspaper, social media) and train on something similar, since taggers as a rule rely on genre-specific features that will not generalize well to unseen genres.
That said, you should probably start by taking a look at pretrained models provided by popular NLP libraries. These include the UDPipe models, of which there are four for English. Of the four English corpora used here, GUM is the most genre-diverse. You can use the UDPipe models in Python via spacy-udpipe.
Just use what’s listed here: https://universaldependencies.org/en/index.html
You can also use the XPOS annotation to experiment with the model, these tags are not directly convertible into UD tags.