Fix epsilon_from_gdp to return a valid high-confidence lower bound

[vvv214]

Summary

CanaryScoreAuditor.epsilon_from_gdp should return a high-confidence lower bound. The previous implementation used

np.abs(isf(FPR_ub) - ppf(FNR_ub))

as a shortcut for the reverse D', D direction. With Clopper-Pearson upper confidence bounds, that shortcut can be too optimistic: a negative forward-direction mu is not automatically valid reverse-direction evidence, because the reverse direction needs its own error-rate bounds and threshold frontier.

This PR computes the two directions explicitly:

• one-sided GDP mu on the original D, D' score frontier;
• one-sided GDP mu on the swapped D', D score frontier;
• one Bonferroni budget over both frontiers' FPR/FNR confidence bounds;
• one final conversion from max(mu_forward, mu_reverse, 0) to (epsilon, delta).

Tests

Added coverage for null, small-sample null, forward-separated, and reverse-separated scores. Against the previous implementation:

scenario	previous implementation	this PR
null scores	fails (epsilon ~= 3.54)	passes (epsilon = 0)
small-sample null scores	fails (epsilon ~= 5.92)	passes (epsilon = 0)
forward-separated scores	passes	passes
reverse-separated scores	fails (epsilon = 0)	passes (epsilon ~= 15.81)

Also checked locally:

uvx pyink --check --diff jax_privacy/auditing.py tests/auditing_test.py
git diff --check
python3 -m py_compile jax_privacy/auditing.py tests/auditing_test.py

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[galenandrew-google]

Thank you for your change. The code looks good but I believe the tests can be improved.

[galenandrew-google]

This test is less specific than the earlier test [test_epsilon_from_gdp_tight]. Please remove it.

[galenandrew-google]

If I understand correctly, the change in this PR means that if the in_canary_scores have smaller values on average than out_canary_scores, the returned epsilon is zero. If so, we should test that case like:

in_canary_scores = rng.normal(-1, 1, m)
out_canary_scores = rng.normal(0, 1, m)

I would suggest that instead of this test and test_epsilon_from_gdp_small_sample_null_is_zero below, we have a single parameterized product test that looks at
(large sample, small sample) x (mu 0, mu negative)