feat: add bounded TensorFlow maximum and minimum

[rextio]

Summary

• add exact top-level tf.maximum and tf.minimum lowering for float32 CPU tensors
• support same-rank rank-1/rank-1 and rank-2/rank-2 TensorFlow broadcasting
• reuse the owned same-wheel TFE runtime path and preserve incompatible-shape failures
• certify finite, NaN/Inf, signed-zero, broadcasting, and shape-error behavior
• document the supported and excluded extrema surface

Scope

This remains CPU-only native AOT. Mixed-rank extrema, scalar operands, aliases,
raw ops, CUDA placement, and all other TensorFlow CUDA work remain excluded.

Verification

• pytest: 337 passed, 1 skipped
• ruff check src tests
• mypy