Add div op workaround

[vladimirjovanovicTT]

Broadcast of non-commutative ops remains to be an issue in metal:
tenstorrent/tt-metal#14852

We are encountering this with div op in tt-xla:
tenstorrent/tt-xla#288

Since the op appears in backward pass of softmax op, this is essentially blocking us from training any classifier network.
In FFE, div op is being decomposed into reciprocal+multiply, which seems like a good solution here as well.
I suggest to introduce a workaround pass for this, where div is decomposed into reciprocal+multiply.

Some historical context:
#842

[vladimirjovanovicTT]

Training crew would like to work on this; however we will try to push for the solution on metal first. Once we get an ETA from them, we will have a clearer picture on how/whether to continue.

[sdjordjevicTT]

Training crew would like to work on this; however we will try to push for the solution on metal first. Once we get an ETA from them, we will have a clearer picture on how/whether to continue.

I will add here for the full context, that we are now pushing on the metal side this one:
tenstorrent/tt-metal#12798