Validate DiscreteArray maxima against integer dtype limits

[sylvesterkaczmarek]

Fixes #22.

Change

Compare num_values - 1 with np.iinfo(dtype).max instead of comparing NumPy dtypes. The previous check could accept a maximum outside a signed integer's range. With NumPy 1.26 this can construct an invalid spec, such as DiscreteArray(129, np.int8) with maximum -128; with NumPy 2.5 it reaches a cast and raises OverflowError instead of the documented diagnostic.

This is a one-line validation correction. Valid signed/unsigned ranges, including their exact upper endpoints, remain accepted. Public signatures, generated values, replacement and serialization behavior are otherwise unchanged.

Verification

All 38 new regression/control cases pass. They cover all eight integer dtypes, exact endpoints, out-of-range counts, dtype strings/objects, NumPy integer counts, replacement and pickle round trips. Original production code fails 14 cases and passes 24 controls.

The complete repository suite passes 202 tests and six subtests locally on macOS/Python 3.12 and in exact-commit hosted validation on Ubuntu with Python 3.10/NumPy 1.26.4 and Python 3.12/NumPy 2.5.3. The hosted jobs test 3335013894f56fb598ef49ad60753c5ddde9fedf, reproduce the original failures, pass the original existing suite, restore the submitted bytes and rerun the regressions.

Source-distribution and wheel builds pass. All 38 cases also pass against installed wheels outside the checkout with the import location verified. New-test formatting, lint, dependency and patch checks pass. Logs, JUnit reports, coverage and packages are workflow artifacts; the validation workflow stays on a separate fork-only branch.

Run the regression module with python -m pytest dm_env/discrete_array_bounds_test.py.

Only the specification implementation and new test module change. No production dependencies, upstream workflows or existing test expectations are modified.