Add a coverage job and badge

[donald-heddesheimer]

The CI is the strongest thing in this repo and nothing surfaced it. This adds a coverage job and a badge, plus a paragraph in the README that says out loud what the matrix already does.

The scoping problem solves itself

CADENCE_HAS_CUDA comes from __has_include(<cuda_runtime.h>) in platform.h, so on a runner with no CUDA toolkit every device path is preprocessed away before gcov sees it. That means no gcovr filter gymnastics and, more importantly, no dishonest denominator: the figure is what the no-GPU tests genuinely reach, not one diluted by lines only a GPU could execute. Those lines are compile-checked by the cuda job in the same run and exercised by cadence_device_tests on real hardware.

The number, measured in advance

Run locally against main with the same configuration the job uses:

include/cadence/cadence.h                      8        8   100%
include/cadence/detail/config.h               46       33    71%
include/cadence/detail/labels.h               20       20   100%
include/cadence/detail/nvtx.h                  4        4   100%
include/cadence/detail/registry.h            228      222    97%
include/cadence/detail/report.h              216      173    80%
include/cadence/detail/scopes.h               15       15   100%
include/cadence/detail/stats.h                66       65    98%
include/cadence/detail/thread_state.h          4        4   100%
include/cadence/detail/trace.h                62       54    87%
TOTAL                                        669      598    89%

89.4% lines, 98.7% functions. The two soft spots are honest ones: config.h is environment-override parsing that a test would have to setenv its way through, and report.h's gap is almost entirely WriteSummary, which #8 covers.

The gate is set below the number, deliberately

--fail-under-line 85 against a current 89.4%. A threshold pinned to the high-water mark turns every honest refactor into a coverage failure; this job is here to catch a regression, not to make the number a ratchet. The gate is what fails a bad change — Codecov is presentation, so fail_ci_if_error: false keeps a third-party outage from turning the job red.

The number is also written to $GITHUB_STEP_SUMMARY, so it is readable on the run page without opening a log or depending on Codecov at all.

Needs you before merging

1. Authorize the Codecov app on this repo.
2. Add CODECOV_TOKEN to Actions secrets.

Until then the job runs and gates correctly, but the README badge will read "unknown" rather than a percentage. Worth holding the merge until the token is in, or merging now and accepting a grey badge for a day — your call.