Add TrOCR example

[robertknight]

This is a vision-to-text example with a very similar structure to the DistilViT example.

Different TrOCR model sizes use different tokenizers. This example works with the base model which uses a BPE tokenizer, but not the small model which uses a unigram tokenizer.

Compared to Ocrs the models are much larger and thus slower to execute. However being a bigger model it also has more capacity.

TODO:

• Implement the If operator. This will allow using the "merged" output model from Optimum (decoder_model_merged.onnx), which is faster than using the cache-less model (decoder_model.onnx) alone and more size-efficient than using separate models for the initial run and subsequent runs. See Implement If operator #306.
• Support cross-attention KV-caches in rten-generate. These are the past_key_values.{layer}.encoder.{key,value} inputs that Optimum uses. Unlike self-attention KV-caches these are generated once when the encoder is run for the first time and skipped in subsequent runs (Support cross-attention key-value caches in rten-generate #318)
• Investigate why LayerNormalization op is not fused in the decoder
  • The problem is that the "shift and scale" pattern in fuse_layer_norm doesn't match because it expects arguments to the Add and Mul operators to be constants. However they are actually value nodes which capture values from constants defined in the parent graph.