Coverage task does not maintain internal invariant on error

[ranweiler]

OneFuzz: 2.23.1
OS: Windows

In the coverage task, the (blocking, non-async) recording itself occurs on a new OS thread, via spawn_blocking().

To avoid needing to use sync primitives, we just Option::take() the TaskContext's ModuleCache, and pass it to the worker thread, restoring it on completion. However, if recording fails, we now retry some number of times, as of #978. But recording can fail while the worker thread owns the ModuleCache, which it will drop on its return with Err. We will then early-exit TaskContext::record_impl(), without restoring the Option<ModuleCache>, which can result in a panic.

Switch to using threadsafe smart pointers, instead.