Don't mmap datasets

[kousu]

nibabel is sort of convoluted, but you can see here it has this mmap parameter:

        mmap : {True, False, 'c', 'r'}, optional, keyword only
            `mmap` controls the use of numpy memory mapping for reading image
            array data.  If False, do not try numpy ``memmap`` for data array.
            If one of {'c', 'r'}, try numpy memmap with ``mode=mmap``.  A
            `mmap` value of True gives the same behavior as ``mmap='c'``.  If
            image data file cannot be memory-mapped, ignore `mmap` value and
            read array from file.)

It looks like they turned it on by default, with the assumption that it was a performance boost, but I think on the NVMe drive on romane it might be holding us back by triggering thousands of tiny I/Os in parallel. The NVMe drive apparently can handle 400MB/s (maybe more), but not when we force it to give us individual values in the order we ask for them, which I suspect mmap() is probably implicitly doing.

https://stackoverflow.com/questions/9817233/why-mmap-is-faster-than-sequential-io

It can be - there are pros and cons, listed below. When you really have reason to care, always benchmark both.

This is all a big guess but it is an attempt to fix https://github.com/neuropoly/computers/issues/371.

I don't know if this should become a permanent part of ivadomed, but I would like you to test to see if it improves performance on our hardware.