Add support for PaddleOCRv5 models for character recognition.

[tjanczak]

Description

Add converters for PaddleOCR v5 recognition models; add support to download models from HuggingFace repository + load vocabulary form HF config file.

Fixes # (issue)

Any Newly Introduced Dependencies

N/A.

How Has This Been Tested?

Validated locally; to be added to CI tests as part of new sample.

Checklist:

• I agree to use the MIT license for my code changes.
• I have not introduced any 3rd party components incompatible with MIT.
• I have not included any company confidential information, trade secret, password or security token.
• I have performed a self-review of my code.

[oonyshch]

I see this line computes log(1/exp_sum + 1e-10) as a log-softmax approximation. the 1e-10 is added after the division to prevent log(0), but this is numerically odd: standard log-softmax would be log_max - log(exp_sum), not this form.
please check for correctness

[tjanczak]

this is to avoid divide by zero error, changed to explicit check if not zero

[oonyshch]

MODEL_DIR should be declared as local
the other variables in this function (MODEL_NAME, DST_FILE1, DST_FILE2) all use local, but MODEL_DIR is in the parent scope

[oonyshch]

these lines could set vocab_size and seq_len to 0 if the tensor rank is wrong, but the subsequent check at line 224 produces a misleading error message "Unexpected vocabulary size".
It would be clearer to add an explicit tensor rank check, that would give a clear error when the model output shape is wrong, rather than defaulting and failing later

[oonyshch]

So far everything that I've not mentioned in comments LGTM