Implicit broadcast within divide op is blocking MNIST training

[umalesTT]

Backward pass for MLP training for MNIST with cross-entropy loss fails with error: BinaryOpType cannot be mapped to BcastOpMath
This is due to metal issue tenstorrent/tt-metal#15216 and, on a more general side, issue that binary eltwise ops support broadcasting only for second input, which fails for non-commutative ops (as explained in tenstorrent/tt-metal#14852).
Concretely, in the case of this backward pass, broadcast in division is needed for the first input (from 1x1xf32 to 4x1xf32) as may be seen below.

%72 = "ttnn.div"(%71, %70) : (tensor<1x1xf32, #ttnn.ttnn_layout<(d0, d1) -> (d0, d1), <1x1>, memref<1x1x!tt.tile<32x32, f32>, #ttnn.buffer_type<dram>>, <interleaved>>>, tensor<4x1xf32, #ttnn.ttnn_layout<(d0, d1) -> (d0, d1), <1x1>, memref<1x1x!tt.tile<32x32, f32>, #ttnn.buffer_type<dram>>, <interleaved>>>) -> tensor<4x1xf32, #ttnn.ttnn_layout<(d0, d1) -> (d0, d1), <1x1>, memref<1x1x!tt.tile<32x32, f32>, #ttnn.buffer_type<dram>>, <interleaved>>>

Cases like this should be covered in https://github.com/tenstorrent/tt-xla/blob/main/tests/jax/ops/test_divide.py, as well as in tests for other binary eltwise non-commutative ops.