Add approx_rational for rational approximations of Computables

[TimTheBig]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces the approx_rational method to Computable in src/computable/node.rs, which provides a dyadic rational approximation within a specified precision. It includes a fast-path optimization for exact-rational leaves with power-of-two denominators and adds comprehensive unit tests. The feedback suggests simplifying the fast-path check by reusing the existing dyadic_denominator_shift method on Rational to improve code maintainability.

[timschmidt]

What is the goal here?

Real::exact_rational_or_approx* ignores Real.rational for non-One classes. In src/real/arithmetic.rs:1352 and src/real/arithmetic.rs:1385, the code approximates only the stored/computed class payload. But Real is rational * class, and many symbolic constants are scaled this way, for example Real::tau() is 2 * Pi with a Computable::pi() payload. The PR returns 201/64 for Real::tau().exact_rational_or_approx(-6), but the correct result is 201/32. Same bug affects borrowed form, pi/2, scaled logs/sqrts, and generic symbolic scaled values.

Real.fold() already manages approximation at the level of Real, which is what you need to correctly address the above issue.

[TimTheBig]

What is the goal here?

The goal is to provide both convenience and correctness. I could see others making this function so why not make an internal one that takes advantage of shortcuts.

As for why you would need it, sometimes you need a rational-only operation that an f64 would be to imprecise or small for.