Answering my own question after some more googling and doc search. I am not 100% sure but I think the right answer is: no, this is not possible with a parfor construct. The ease of use of parfor comes with some limitations and this is one of them. For sharing a large array between workers the code would need to be rewritten as a spmd region and use the distributed function. I expect this might not be ideal for small-to-medium arrays though, that we would not want fragmented across memory, but possibly ML is smart enough to optimize this behavior behind the scenes.
As a last resort the sharedmatrix fx submission suggested above might work but it sounds pretty risky. A lot of talk of segmentation faults in the comments ;)