[infer] Add progress_callback to inference, judge, and synthesis APIs

[rlehman221]

Summary

• Adds an optional progress_callback: Callable[[int, int], None] parameter to BaseInferenceEngine.infer(), BaseJudge.judge(), and AttributeSynthesizer.synthesize()
• The callback fires (completed_count, total_count) after each item finishes, enabling callers to report granular progress during bulk operations
• RemoteInferenceEngine fires the callback per async task completion in the gather loop; other engines accept the parameter but don't use it yet

Motivation

When the Oumi Enterprise worker calls bulk inference/judge/synthesis, it passes all rows at once. Without this callback, the worker can only update heartbeat progress at the start and end of the call — leaving users with no progress updates for potentially long-running operations.

Changes

File	Change
base_inference_engine.py	progress_callback on infer() and _infer_online()
remote_inference_engine.py	Fires callback per task in async gather loop
native_text_inference_engine.py	Accepts param in _infer_online()
vllm_inference_engine.py	Accepts param in _infer_online()
llama_cpp_inference_engine.py	Accepts param in _infer_online()
base_judge.py	progress_callback on judge() and _infer(), forwarded to engine
attribute_synthesizer.py	progress_callback on synthesize(), forwarded to engine

Design

• Callback is optional (None default) — fully backwards compatible
• Thread-safe: asyncio.gather() runs on a single event loop thread, so the nonlocal counter won't race
• Best-effort: callback exceptions are silently caught to never crash inference
• Engine-level: judge and synthesis just forward the callback to the inference engine