Description

1. Fixes an ungracefully unhandled type case (in materialising a symbolic dimension), which always caused errors when a map-typed operator was used.
2. Completes ZipMap type inference to properly type the map value types as scalar float tensors.
3. Adds a test for running ZipMap type inference (and hence testing that maps now are in fact supported by shape inference).

Motivation and Context

ONNX support for the map type is poor and the least documented. This is a small step towards adding more operators using map for more advanced pre-processing capabilities, more dynamic than what is allowed by the existing, attribute-oriented, ai.onnx.ml operators (like LabelEncoder).

Currently, any and all uses of shape inference where a map type is present results in the following error (e.g. for ZipMap):

onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:ZipMap): [ShapeInferenceError] type case unsupported for symbolic shape inference. inferred=5

This is due to a piece of code unrelated to a specific inference implementation, and hence map operators cannot be used at all as they will always raise this error. The shape inference test module had a similar lack of a map case. I skimmed through the code and I didn't see other such cases.

Also, related comment where I uncovered this - skl2onnx-produced classifiers use ZipMap by default and hence can never succeed in shape inference.