Fixes for fp16 tensors and datatypes in ONNX

[maltanar]

QONNX traditionally made a number of assumptions about how quantization is represented: using float32 carrier datatypes, with quantization introduced through explicit Quant nodes and optionally with datatype annotations on tensors (InferDataTypes). Some of these baked-in assumptions cause inflexibility around other datatypes. This PR relaxes a few of these restrictions around float16 by introducing the following changes:

• ModelWrapper.get_tensordatatype previously always returned FLOAT32 for any tensor without a datatype annotation. Now an additional check is performed: if there is no datatype annotation, we look at the underlying container datatype to have the option of returning FLOAT16 instead.
• InferDataTypes supports preserving FLOAT16 either from annotations or from the container datatype.
• gen_finn_dt_tensor now returns np.float16 if the desired data type is specified to be FLOAT16.
• Quant nodes will preserve float16 tensors if those were originally provided as input/outputs.