Update num_proc handling for vllm and Ray mode

[ArdalanM]

Summary

TextTaggingByPromptMapper was unconditionally setting self.num_proc = 1, which in a
Ray+vLLM run capped the actor pool to a single GPU. The only available multi-GPU strategy
was tensor parallelism (tensor_parallel_size=N); data parallelism was silently broken.

This fix makes num_proc = 1 conditional on the execution backend, so Ray can now schedule
one vLLM actor per GPU — enabling both data parallelism, tensor parallelism, or a
combination of both.

[gemini-code-assist]

Code Review

This pull request modifies the initialization of the TextTaggingByPromptMapper to conditionally set the number of processes based on whether vLLM and Ray mode are enabled. The goal is to allow Ray to manage data parallelism by scheduling actors per GPU. A review comment suggests that the current implementation unnecessarily restricts parallelism for the HuggingFace backend when running on Ray and provides a suggestion to simplify the logic to check only for Ray mode.