Reject NaN RZZ angles during primitive validation

[sankalpsthakur]

Summary

Reject fixed and parameterized RZZ angles whose evaluated value is NaN. This keeps invalid input inside the existing IBMInputValueError validation path instead of accepting a fixed NaN or leaking a raw Qiskit binding error.

Details and comments

• evaluate parameter expressions with bind_all before validating the resulting angle
• preserve valid expressions where a NaN-valued parameter is algebraically eliminated
• extend fixed and parameterized RZZ regression coverage, including the eliminated-NaN case
• add the required Towncrier bug-fix fragment

Validation:

• python -m pytest -q test/unit/test_sampler.py: 43 passed, 3 subtests passed
• ruff check on touched Python files: passed
• ruff format --check on touched Python files: passed
• mypy --no-incremental on touched Python files: passed
• git diff --check: passed
• additional automated code review: no findings after the eliminated-parameter revision

Why this matters

NaN gate angles should fail fast at validation time with a clear error, not propagate into backend execution or produce undefined results on real hardware.

Related

Fixes #2452

AI/LLM disclosure

• I didn't use LLM tooling, or only used it privately.
• I used OpenAI Codex to help write this PR description.
• I used OpenAI Codex to generate or modify code and tests.

[CLAassistant]

All committers have signed the CLA.