CAGRA tech debt: distance descriptor and workspace memory

[achirkin]

This PR introduces two changes:

1. Refactor dataset_descriptor_host to pass and cache it by value while keeping the state in a thread-safe object in a shared pointers. Before this, the descriptor host itself was kept in shared pointer in LRU cache and was passed by reference; as a result, it could in theory die due to cache eviction while still being used via references to it.
2. Adjust the temporary buffers to always use the workspace resource in all CAGRA algo implementations (as of now, only SINGLE_CTA algo does this; the PR expands the change to MULTI_CTA and MULTI_KERNEL).

Both of the changes are required for effective use of stream-ordered dynamic batching #261 (1. fixes crashes and 2. fixes thread-blocking behavior).