Fix math reduction conversion and numeric protocol handling - #9666

Summary

Align math reduction protocol handling with CPython 3.15:

• Materialize both dist points and check dimensions before converting coordinates; convert matching coordinates in p/q order and read tuple subclasses directly.
• Restrict prod fast paths to exact builtin types, preserving float subclass multiplication and empty-product results. Simplify the staged accumulator and enable its integer path for explicit integer starts.
• Preserve sumprod iterator construction and exception order, including errors paired with exhaustion. Include bool weights in compensated float accumulation.

Uses existing typed argument binding, sequence conversion, arithmetic, and pymath norm/compensation machinery. No tests, Python files, or workflows changed.

Source comparison

Compared against CPython revision d220409a13b77de7a6c30abe16ea90db01528edc:

• math_dist_impl
• math_sumprod_impl
• math_prod_impl

Validation

• Passed rustfmt and git diff --check.
• Passed 12 bounded CPython 3.15.0 reference probes covering dist conversion and dimension-check order, tuple subclasses, prod subclass dispatch and empty results, sumprod iterator errors/construction order, and compensated cancellation with bool weights in both orientations.
• Passed cargo check and strict Clippy (-D warnings) for rustpython-vm, rustpython-host_env, and rustpython-stdlib with rustpython-stdlib/threading, locked/offline and one build job, against base 0b081b0.
• Workspace/C-API tests and RustPython test_math execution were not run. Reference probes establish expected CPython behavior, not RustPython runtime correctness. Native RustPython execution is held pending the required safety guard.

This does not claim full math parity. In particular, sumprod retains its existing O(n) buffered float-pair storage rather than implementing CPython’s streaming accumulator.

AI-assisted implementation and source comparison using OpenAI Codex (model version not exposed), directed by the maintainer.