Make cgf_diagnostics::teardown() a no-op instead of resetting the instance

[GagaLP]

After a while of debugging a strange Access Violation error on Windows during teardown of Celerity, I traced it back to cgf_diagnostics::teardown().

cgf_diagnostics stores its singleton in a thread_local std::unique_ptr. teardown() is called from runtime::impl's shutdown path, which runs when runtime::s_instance (a plain static member) is destroyed, and if the application never calls runtime::shutdown() explicitly, that happens as part of normal static-object teardown after main() returns.

The C++ standard leaves the relative order of static-object destruction and thread_local destruction for the main thread undefined. On Windows, these use separate mechanisms (the CRT's atexit/exit machinery vs. TLS callbacks), and in this case the thread_local storage for the main thread was already gone by the time teardown() called m_instance.reset() on it, resulting in access to unavailable memory.

Removing the m_instance.reset() call fixes this issue while keeping the logic the same: the whole point of using a smart pointer here is that its cleanup is already handled automatically, so an explicit reset in teardown() is redundant.

[github-actions]

Check-perf-impact results: (ae6918621b46271c2f10d6eb978fe95d)

❓ No new benchmark data submitted. ❓
Please re-run the microbenchmarks and include the results if your commit could potentially affect performance.