flobee.at.gmail.dot.com shared "readfile_chunked" function. It does work, but you may encounter memory exhaustion using "fread". Meanwhile "stream_copy_to_stream" seems to utilize the same amount of memory as "readfile". At least, when I was testing "download" function for my https://github.com/Simbiat/HTTP20 library on 1.5G file with 256M memory limitation that was the case: "fread" I got peak memory usage of ~240M, while with "stream_copy_to_stream" - ~150M.
It does not mean that you can fully escape memory exhaustion, though: if you are reading too much at a time, you can still encounter it. That is why in my library I use a helper function ("speedLimit") to calculate whether selected speed limit will fit the available memory (while allowing some headroom).
You can read comments in the code itself for more details and raise issues for the library, if you think something is incorrect there (especially since it's WIP at the moment of writing this), but so far I am able to get consistent behavior with it.