feat: add comprehensive API and infrastructure cost tracking (#192)

Add a centralized, thread-safe CostTracker that records token usage
from every VLM/LLM API call and infrastructure time (GPU/VM hours).

The tracker is integrated at the vlm_call() level so all 15+ callers
automatically get cost tracking without any changes. Key integration
points have cost_label tags for per-component breakdown (planner,
grounder, vlm_judge, demo_verify, etc.).

- New openadapt_evals/cost_tracker.py with global singleton, pricing
  tables, JSON persistence, and human-readable summary output
- vlm.py extracts response.usage tokens from both OpenAI and Anthropic
  responses and reports to the tracker
- 18 unit tests covering pricing lookup, aggregation, thread safety,
  persistence, and vlm.py integration

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: abrichr

openadapt_evals/agents/demo_guided_agent.py

@@ -315,6 +315,7 @@ def verify_step(
                 model=self._verify_model,
                 provider=self._verify_provider,
                 max_tokens=256,
+                cost_label="demo_verify",
             )
 
             parsed = extract_json(raw)

openadapt_evals/agents/planner_grounder_agent.py

@@ -670,6 +670,7 @@ def _call_planner(
             model=self._planner,
             provider=self._planner_provider,
             max_tokens=512,
+            cost_label="planner",
         )
 
         logger.debug("Planner raw output: %s", raw[:500])
@@ -733,6 +734,7 @@ def _call_grounder(
             model=self._grounder,
             provider=self._grounder_provider,
             max_tokens=256,
+            cost_label="grounder",
         )
 
         logger.debug("Grounder raw output: %s", raw[:500])
@@ -756,6 +758,7 @@ def _call_grounder(
                 model=self._grounder,
                 provider=self._grounder_provider,
                 max_tokens=128,
+                cost_label="grounder_retry",
             )
             action = parse_action_json(raw2)
             if action.type != "done":

openadapt_evals/correction_parser.py

@@ -56,6 +56,7 @@ def parse_correction(
         model=model,
         provider=provider,
         max_tokens=512,
+        cost_label="correction_parser",
     )
 
     # Extract JSON from response