Support pluggable workflow caches

There are cases where the current integrated LRU cache for workflow threads may not sufficiently capture the optimal strategy for all workloads.

For example, some high-volume workloads allow the cache to fill with workflows that are effectively abandoned because a remote worker completed them without any invalidation signal to clear the defunct task within the current worker. These abandoned workflows add extra eviction overhead for new, unrelated workflows, and the system may sometimes benefit from alternate eviction strategies that differ from the current synchronous LRU algorithm. Examples include augmenting the LRU with a TTL model, or adding an asynchronous/background eviction mechanism.

The proposal is to add a mechanism allowing a pluggable cache that defaults to the current LRU design, allowing SDK end-users to supply their own implementations when desired.

Start by locating the current integrated LRU cache for workflow threads and reviewing how workflows enter, remain in, and leave it. Define the pluggable cache boundary, the default LRU behavior, and the user-supplied implementation contract; done means alternate strategies can be supplied without changing existing default behavior.