Add random sampling to dm_env array specifications

[sylvesterkaczmarek]

Summary

Addresses #14 by adding sample(rng=None) to dm_env array specs for random values consistent with each spec, analogous to sampling Gym spaces. generate_value() remains unchanged and deterministic.

• Array.sample(): dtype-aware random sampling for boolean, integer (including full uint64 range), floating and complex arrays, at scalar or multidimensional shapes.
• BoundedArray.sample(): uniformly sample integer bounds inclusively, with scalar or elementwise broadcasting; uniformly sample finite floating bounds, and sample valid values when one or both bounds are infinite.
• DiscreteArray inherits inclusive integer sampling, producing a valid action index.
• StringArray.sample() generates short random ASCII str or bytes strings in the declared shape.
• Accept an optional numpy.random.Generator to make sample sequences reproducible; without one, create a new generator. Reject unsupported nonnumeric plain Array dtypes rather than silently returning invalid values.
• Document usage, distributions and the existing generate_value() behavior in the API guide.

Verification

• 185 tests and 17 subtests passed across the entire dm_env test directory, including the new randomized shape, dtype, bound, seeded reproducibility, discrete, empty-shape, string-type and infinity edge cases.
• Checked Python compilation, focused Ruff lint, Pyink formatting and Git whitespace checks.
• Tested on Python 3.12 and NumPy 2.5.3, without environment or network requests. Other Python/NumPy combinations have not been validated locally.

This is an additive API; no existing sampling or environment behavior is replaced.