Validate structural arguments (num_bands, truncated sizes) in accounting events

[r1cksync]

Bug

Several jax_privacy.accounting event builders forward user-controlled structural arguments into divisions / sizing without validating them, so invalid inputs either crash opaquely or are silently accepted:

>>> from jax_privacy import accounting
>>> accounting.amplified_bandmf_event(1.0, 128, num_bands=0, sampling_prob=0.01)
ZeroDivisionError: division by zero
>>> accounting.amplified_bandmf_event(1.0, 128, num_bands=-4, sampling_prob=0.01)
ValueError: Expected iterations=-32 >= 0.          # misleading: blames iterations
>>> accounting.truncated_dpsgd_event(1.0, 10, sampling_prob=0.1,
...     num_examples=-5, truncated_batch_size=-3)
<SelfComposedDpEvent ...>                           # silently accepted, nonsensical

num_bands flows into math.ceil(iterations / num_bands) in amplified_bandmf_event and truncated_amplified_bandmf_event; num_examples / truncated_batch_size flow into TruncatedSubsampledGaussianDpEvent in truncated_dpsgd_event. None were validated, while sibling builders (fixed_dpsgd_event via _validate_fixed_args) already validate their sizes.

Fix

• _validate.positive(num_bands=num_bands) in both BandMF builders (must be ≥ 1 to partition the data and to divide by).
• _validate.non_negative(num_examples=..., truncated_batch_size=...) in truncated_dpsgd_event, matching the _validate_fixed_args convention for dataset/batch sizes.

Only previously-invalid inputs (non-positive num_bands, negative sizes) are rejected — valid inputs are unchanged. This mirrors the already-merged #255 and the existing _validate pattern.

Testing

python -m pytest tests/accounting_test.py — 36 passed. The new parameterized tests cover each newly-guarded argument × invalid value, plus positive regressions; they fail (8 failures + 2 ZeroDivisionError) against the pre-fix code and pass after.

[github-actions]

This PR has been idle for 7 days. Please provide an update or review.

[arung54]

Thanks for the PR. No issues with the changes in logic, but a few comments to keep the style consistent. Will be happy to approve once they're addressed.

[arung54]

I think '_unchanged' here and in the below test names is unnecessary and if anything adds confusion.

[arung54]

I think none of the comments introduced in this CL are necessary, and would suggest removing them all.

From the test names I think it is pretty easy to infer what each test is supposed to do, and the logic can be inferred from reading the file being tested. We generally prefer not to include comments if they only explain something that can be inferred by someone with knowledge of Python.