feat(checks): aggregate LLM token usage on SuiteResult

[davidberenstein1957]

Feature Request

Roll up token usage from LLM-backed steps (judges, attackers, input generators) onto SuiteResult, and print a one-line summary in the suite report when totals are non-zero.

Motivation

giskard-llm already returns Usage on completions. Suite/scan reports do not expose aggregates, so “how many tokens did this cost?” needs custom instrumentation.

Implementation plan

1. Define a small aggregate model (reuse or wrap giskard.llm Usage fields):

   class SuiteUsage(BaseModel):
       prompt_tokens: int = 0
       completion_tokens: int = 0
       total_tokens: int = 0

2. Collect during run (preferred): mutable collector on run context incremented wherever acompletion / generator usage is known. Fallback: sum from check details if usage is already stored there.

3. Attach to SuiteResult as usage: SuiteUsage (default zeros).

4. Extend print_report() footer when total_tokens > 0.

5. Tests with a fake generator that records known usage → aggregate matches; pure string-check suites stay at zero / omit footer line.

API usage

from giskard.checks import Scenario, Suite, Equals

suite = Suite().append(
    Scenario("no-llm")
    .interact(inputs="hi", outputs="hi")
    .check(Equals(expected="hi"))  # after key-default issue; or key=...
)
result = await suite.run()
assert result.usage.total_tokens == 0

# After an LLM-judge / scan run:
result = await vulnerability_scan(target=my_agent, description="...", languages=["en"])
print(result.usage.prompt_tokens, result.usage.completion_tokens, result.usage.total_tokens)

# Programmatic gate (optional user pattern — not built-in budget)
if result.usage.total_tokens > 500_000:
    raise RuntimeError("eval exceeded token budget")

Example report footer

SUMMARY
  36 scenarios  pass=30  fail=5  error=1  pass_rate=83.3%
  tokens: in=128400  out=39200  total=167600

Example JSON (suite dump fragment)

{
  "usage": {
    "prompt_tokens": 128400,
    "completion_tokens": 39200,
    "total_tokens": 167600
  }
}

Dollars / pricing maps are out of scope for v1.

Acceptance criteria

• LLM-backed runs expose aggregated token counts on SuiteResult when usage was recorded
• Report includes a one-line usage summary when total_tokens > 0
• Unit test: fake recorded usage → aggregate matches
• Non-LLM suites: zeros and no misleading footer

Out of scope

• Hard spend budgets / early-stop
• Multi-provider $ tables
• Generation progress UI (separate)

[linear-code]

OSS-108