Fix pea.sparse on IPU

[DouglasOrr]

Addresses two issues:

1. The example script breaks with No default opset version defined for domain 'ai.graphcore.pea'. This seems to be happening since the 3.2 update, and does not occur in tests. Running under gdb, seems to be triggered by shape inference and the non-extensibility of PopART's Ir::getDefaultOpsetVersion. I'm not sure if this might hit other custom ops using the "ai.graphcore.pea" domain, or if there is a better workaround than just reverting to "ai.graphcore".
2. A known issue - since spmm does not add its' attributes under appendAttributes/appendOutlineAttributes, we get spurious cache hits "every SpMM is the same".

---

Full error:

2023-04-24T11:03:37.971078Z popart:popart 50817.50817 E: No default opset version defined for domain 'ai.graphcore.pea'

[0] popart::OpManager::createOpInGraph(onnx::NodeProto const&, popart::Graph&)
[1] popart::Graph::constructFromOnnxGraph(onnx::GraphProto const&)

[DouglasOrr]

(Multiple reviewers - just for whoever gets there first; cheers!)

[AlCatt91]

I believe I once came across the No default opset version defined for domain 'ai.graphcore.pea' error in 3.2 using the distance_matrix custom op, however at the time I dismissed it since there were also other bugs involved in my script and I thought it was a consequence of them... However, from what you report it might actually be an issue: I think at this point it might be safer to use the ai.graphcore domain everywhere in PEA.

[DouglasOrr]

Good point, thanks for the heads up! I've pushed a change to use ai.graphcore everywhere (there should be no downsides to doing so, as a name collision is unlikely, and could be worked around in any case.)

[AlCatt91]

Thanks for updating the other ops too, I agree this should not cause major issues in terms of collisions. Looks good to me!