Make LazyInit initialization single-flight

[adenzhou1350]

Summary

• make concurrent first use of LazyInit initialize exactly one shared object
• preserve retry after factory exceptions or null results
• keep copies on the same initialization state and report moved-from use cleanly

Problem

LazyInit::get() currently reads and writes the same std::shared_ptr without synchronization. Concurrent first use is therefore a data race and may also run the factory more than once. This matters for lazily constructed runtimes when several Python or host threads enter the extension together.

Validation

• Upstream baseline: a synchronized 16-thread first-use test reports 16 factory calls instead of one.
• Candidate: 25 rounds x 16 threads each produced one factory call, one shared address, and the expected value.
• ThreadSanitizer: the same standalone probe reports a data race in baseline LazyInit<int>::operator->(); the candidate exits cleanly.
• Exception retry, empty factory, null result, assignment, copy, move, -Wall -Wextra -Werror, Python syntax compilation, and git diff --check pass.

This is host-side correctness work; it does not claim a GPU performance improvement.