tt-metal: Untilize sweep test does not work for multicore (last shape)

test: tests/tt_eager/python_api_testing/sweep_tests/pytests/tt_dnn/test_untilize.py::test_untilize_test[untilize_args0-input_shapes2]

This needs to be fixed for multicore, and then change the default untilize to use multicore once this is fixed.

About this issue

On Grayskull, this test cant run cause it runs out of L1. Couple of things I have noticed with untilize:

Single-core and multi-core both take full width per core. Single-core version properly figures out how much of the width to buffer, while multi-core naively double buffers on the full width. That’s causing this test to run out of memory.
If each core gets the full width and at least tile HEIGHT, I would expect multi-core to just work as well.

What’s needed to enable multi-core by default is:

defining some shapes to test before we call it working. So far, I haven’t seen a case fail on Grayskull multi-core.
clean up multi_core and pack_untilize flags. We should just use them both as True and fix what’s not working. It seems like it’s just a matter of getting pack_untilize to work with longer widths.