gh-70450: Fix MSVC /fp:strict builds

[yhay81]

Fixes #70450.

Summary

• replace floating-point arithmetic in static initializers with equivalent binary64 literals
• keep the cmath special-value tables read-only and decode their finite NaN sentinel when a value is read
• create the exported math and cmath NaN constants with nan(""), avoiding floating-point exceptions during strict-mode imports
• document the precomputed literals and add exact-value regression tests
• add a build NEWS entry

Root cause

MSVC's strict floating-point mode does not accept the affected arithmetic expressions as static initializers. In C mode, the UCRT NAN expansion also performs runtime floating-point arithmetic under /fp:strict; using it during module initialization set FE_INVALID | FE_OVERFLOW | FE_INEXACT.

The cmath tables therefore retain a finite sentinel in static storage and convert only the selected result to a quiet NaN with nan(""). This keeps the tables const, avoids process-global initialization state, and preserves two-dimensional array bounds.

Validation

• built PCbuild/pythoncore.vcxproj in Debug x64 with ICCBuild=true (selecting FloatingPointModel=Strict): passed
• built the same project with the normal floating-point configuration: passed
• cleared the UCRT floating-point status in fresh processes, then imported math and cmath: both remained 0x0, with errno 0
• ran python_d.exe -m test --timeout 300 -j1 test_math test_cmath test_float test_strtod: 186 tests passed, 4 skipped
• ran Tools/c-analyzer/check-c-globals.py --format summary --traceback Modules/cmathmodule.c with the GCC preprocessor: passed
• ran pre-commit on all six changed files, Tools/patchcheck/patchcheck.py, and git diff --check: passed

AI assistance: This patch was developed with assistance from OpenAI Codex; I reviewed and tested the final changes.

• Issue: Cannot compile with /fp:strict with MSVC #70450

[python-cla-bot]

All commit authors signed the Contributor License Agreement.