Split broadcast so it is always fused in compile - #2318

What this does becomes clear with an example. For the following function

def f(a, b, c):
    bb = mx.broadcast_to(b, c.shape)
    return (a + bb * c), bb.sum()

the compiled graph changes as follows

Benchmarking on transformer training the speedup is consistent but minimal (~3%-4%) across M2, M2 Ultra and A100.

Splitting broadcast is safe because it is in principle a noop, but we could consider splitting some copies (astype) and redoing them if it will allow us to fuse more things.

@awni I tried to think if there is a possibility of messing up the graph but I think it is ok since to split a broadcast means we certainly are going to produce a fused primitive.