Cache CK convolution kernel objects as shared_ptr in Invoker lambda

[amberhassaan]

CK has a DeviceOpFactory::GetInstances() API used by MIOpen Solvers that returns a vector<unique_ptr<T>, where T is an abstraction for a kernel. unique_ptr<T> makes it uncacheable inside MIOpen's Invoker lambda thus preventing reuse and forcing multiple calls to GetInstances() especially at the time of invoking a kernel.

I will put up a PR that promotes unique_ptr<T> to shared_ptr<T> thus allowing us to cacheable inside MIOpen's Invoker.

[atamazov]

[Notice] IIUC this ticket is related to all CK solvers where this optimization is possible (see #2305 (comment))