Support Elu operator, reading axes list from inputs rather than attribute in Reduce* ops

[robertknight]

This addresses a few issues I found while playing around with ONNX models from openpilot:

• Support the Elu activation operator
• The latest versions of the Reduce* operators take axes as a dynamic input rather than static attribute.
• Enable the ort-infer.py script to run with models that have float16 inputs

For RTen, the original supercombo.onnx model fails to convert because the weights are in float16 format. However this hack works (at least for getting the model to run, I haven't verified the outputs):

diff --git a/rten-convert/rten_convert/converter.py b/rten-convert/rten_convert/converter.py
index 42b427a6..d4fe5f4e 100644
--- a/rten-convert/rten_convert/converter.py
+++ b/rten-convert/rten_convert/converter.py
@@ -426,6 +426,9 @@ def constant_node_from_onnx_initializer(
         case "float32" | "int32":
             pass
 
+        case "float16":
+            data = data.astype(np.float32)
+
         # Int types that can be widened to int32
         case "bool" | "int8" | "int16":
             data = data.astype(np.int32)

TODO:

• Tests for Elu op
• Tests for Reduce operator change