Correctly round AA conversion to float

[cxzhong]

Fixes #42545.

Problem

AlgebraicReal.__float__ delegated to float(RR(self)). The intermediate 53-bit MPFR value can introduce double rounding, and a lazy algebraic expression that is exactly zero can therefore become a small nonzero Python float. This caused the matrix failure reported in #42545.

Changes

• Convert algebraic interval endpoints through their exact rational values before binary64 rounding.
• Refine the enclosure until both endpoints identify one rounding cell.
• Resolve adjacent-float midpoint and overflow boundaries by exact comparison.
• Preserve representable normal and subnormal values.
• Return a canonical positive zero only when an underflow-sized enclosure still contains zero and its exact sign remains unresolved.
• Update indirect QQbar and LazyAlgebraic float doctests.

This PR contains no matrix or LDLT changes.

Verification

• src/sage/rings/qqbar.py: 1688 doctests passed.
• src/sage/rings/real_lazy.pyx: 286 doctests passed.
• 1010 adversarial rational cases across normal, subnormal, midpoint, signed-zero, and overflow boundaries matched direct rational-to-float conversion.
• The is_positive_semidefinite() throws ZeroDivisionError: division by zero in algebraic field #42545 matrix reproduction now returns True from is_positive_semidefinite().

Checklist

• The title is concise and informative.
• The description explains the change.
• Tests cover the changed behavior and boundary cases.
• Modified doctests pass.

[github-actions]

Documentation preview for this PR (built with commit 52938f8; changes) is ready! 🎉
This preview will update shortly after each push to this PR.