fix(evaluation): reject num_samples=0 in JudgeModelOptions at construction time

[gaurav-gandhi-2411]

🔴 Required Information

Describe the Bug:
JudgeModelOptions.parallelism_limit has ge=1 but its sibling field
num_samples has no lower-bound constraint at all. ge=1 was added to
parallelism_limit by commit 0dbb88c ("feat: parallelize LLM-as-judge
evaluation using asyncio.gather()", 2026-08-19), but that commit did not touch
num_samples. Across this file's entire visible history, num_samples has
never carried a lower bound.

num_samples=0 is never an intentional value anywhere in this codebase — a
repo-wide grep of every num_samples usage in src/ and tests/ finds only
1, 3, or the default 5; nothing ever sets or derives 0. It is, however, a real
live bug surface: LlmAsJudge.evaluate_invocations silently drops any
invocation with zero samples from the result set entirely, with no error and
no NOT_EVALUATED marker. (The "judge model yields no response" exception path
is a separate, already-fixed case on main; this PR is narrower and
orthogonal, targeting only the num_samples=0 misconfiguration itself.)

Steps to Reproduce:
JudgeModelOptions(num_samples=0) # succeeds silently today

Expected Behavior:
Same as JudgeModelOptions(parallelism_limit=0) today — a
pydantic.ValidationError at construction time.

Observed Behavior:
Constructs successfully, then silently drops every invocation evaluated by it
from the aggregated result.

Changes

• src/google/adk/evaluation/eval_metrics.py: add ge=1 to
  JudgeModelOptions.num_samples, matching the existing convention for
  count-like fields in this package — two other examples: run_config.py: max_workers and context_cache_config.py:cache_intervals, both ge=1.
• tests/unittests/evaluation/test_eval_config.py: add
  test_judge_model_options_rejects_zero_num_samples, asserting
  JudgeModelOptions(num_samples=0) raises pydantic.ValidationError.

Note for reviewers: this makes a defensive if not samples: return NOT_EVALUATED branch in simulation/per_turn_user_simulator_quality_v1.py
(~lines 365-369) unreachable via the public constructor — that branch was
never exercised by an intentional num_samples=0, so removing it is optional
cleanup, not required here; left untouched to keep this diff minimal.

Testing

pytest tests/unittests/evaluation/test_eval_config.py -q   → 21 passed
pytest tests/unittests/evaluation/test_llm_as_judge.py -q  → 10 passed

No regressions in either file.

Risk & rollback

Additive validation only — narrows accepted input, does not change behavior
for any value ever actually used (1, 3, 5 across every existing caller/test).
Revert is a single-line removal of ge=1.