[Relay][Frontend][Onnx] GRU Layer Support

[jwfromm]

This PR adds GRU parsing to the onnx frontend. Currently this is done by unrolling the recurrence in a similar way to how we handle LSTMs. Since there's quite a bit of shared code, I've generalized the LSTM converter to an RNN converter and made LSTM and GRU subclasses. For testing, I again replaced the LSTM test function with a more general test_rnn function that supports both LSTM and GRU and should be easily expandable to other RNN functions should we ever want to add any.

[jwfromm]

@masahi, @areusch, @soiferj can you guys take a look at this PR?

[anijain2305]

This might be somewhat relevant - https://discuss.tvm.ai/t/onnx-lstm-op-conversion/7238
Just pointing it out here. I think the GRU might also have the same problems that I encounter with LSTM.

[masahi]

@siju-samuel please help review this PR.

[jwfromm]

I've converted this to a draft as I'll try to incorporate the points made by @anijain2305 in the linked discuss post.

[jwfromm]

I've added a small change to how arguments are parsed in RNNs based on the discussion here: https://discuss.tvm.ai/t/onnx-lstm-op-conversion/7238/5. Our previous implementation assumed that the position of optional arguments could not be known without their name, however, this is not true assuming the nodes are constructing properly. To avoid name based problems, RNNs now use indexing to get inputs. I've updated the onnx_input structure to allow indexing outside of the input bounds but return None in such cases, indicating that an input was not provided.

[masahi]

Thanks @jwfromm @areusch @anijain2305