[skip ci] Fix FITS card float formatting (swev-id: astropy__astropy-14508)

[casey-brooks]

Summary

• tighten _format_float to honor the 20-character limit without losing legacy behavior
• add regression coverage for representative float and complex card values

Testing

• .venv/bin/python -m pytest astropy/io/fits/tests/test_card_float_format.py
• .venv/bin/python -m pytest astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_floating_point_value_card
• .venv/bin/python -m pytest astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_complex_value_card

swev-id: astropy__astropy-14508
#117

[casey-brooks]

Test & Lint Summary

• .venv/bin/python -m pytest astropy/io/fits/tests/test_card_float_format.py ✅
• .venv/bin/python -m pytest astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_floating_point_value_card ✅
• .venv/bin/python -m pytest astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_complex_value_card ✅
• .venv/bin/python -m ruff check astropy/io/fits/tests/test_card_float_format.py ✅

Note: running ruff against the legacy astropy/io/fits/card.py module surfaces numerous pre-existing project warnings; the new helper and tests introduce no additional lint violations.

[noa-lucent]

Thanks for tightening the float formatter and adding regression coverage. I left one blocking issue: the new "just in case" fallback that slices to 20 characters silently produces non-round-tripping tokens instead of surfacing an error when formatting fails. Please replace the fallback with a hard failure (or otherwise make the invariant explicit) so we do not hide formatting bugs.

[casey-brooks]

Test & Lint Summary

• .venv/bin/python -m pytest astropy/io/fits/tests/test_card_float_format.py ✅
• .venv/bin/python -m pytest astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_floating_point_value_card astropy/io/fits/tests/test_header.py::TestHeaderFunctions::test_complex_value_card ✅
• .venv/bin/python -m ruff check astropy/io/fits/tests/test_card_float_format.py ✅

(ruff against the full astropy/io/fits/tests/test_header.py still reports pre-existing project warnings; our edited assertions introduce no new lint findings.)

[noa-lucent]

Thanks for the fast follow-up—the formatter now fails loudly instead of returning truncated tokens, and the new tests cover both the round-tripping success cases and the failure path. Looks good to me.