Fix off-by-one error in vLLM adapter vocab_size calculation

[dmndxld]

Problem

In WhiteboxModelvLLM.post_processing() (src/lm_polygraph/model_adapters/whitebox_model_vllm.py:115), the log_prob
matrix allocation is based on an incorrect vocab_size:

vocab_size = max(
    self.tokenizer.vocab_size, max(self.tokenizer.added_tokens_decoder.keys())
)

added_tokens_decoder.keys() returns token IDs, not a count. When a model has an added token with ID equal to
vocab_size (e.g., Qwen3 has an added token at ID 151668 while vocab_size is 151668), the matrix is too small. This
causes an IndexError at line 122:

log_prob[i, top_tokens] = top_values  # IndexError: index 151668 is out of bounds for dimension 0 with size 151668

Fix

Convert the max added token ID to a size by adding 1:

max_added_token_id = max(self.tokenizer.added_tokens_decoder.keys()) if self.tokenizer.added_tokens_decoder else 0
vocab_size = max(self.tokenizer.vocab_size, max_added_token_id + 1)

Also adds a guard for empty added_tokens_decoder.

Affected models

Any model whose added token IDs >= tokenizer.vocab_size, such as Qwen3.