fix: Run prompt_guard model in a seperate thread

[derekhiggins]

The GPU model usage blocks the CPU. Move it to its own thread. Also wrap in a lock to prevent multiple simultaneous run from exhausting the GPU.

Closes: #1746

What does this PR do?

Runs the prompt_guard model run in its own thread, protected by a lock

Test Plan

tested locally, running multiple runs simultaneous requests, inference was delayed while all safety shields were running with the patch inference is no longer delayed