[CUDA][HIP] Add event pool

[illuhad]

Previously, a new CUDA or HIP event was created for every new operation. Calls to cudaEventCreate() or hipEventCreate can however be costly.

This PR adds an event pool for the CUDA and HIP backends. It works as follows:

• We maintain a per-device list of events that are currently unused (event pool).
• When we need a new event, we first check in the event pool whether there are unused events available
  • if so, we pop one from there and use it.
  • If not, we construct a new one using cudaEventCreate()/hipEventCreate().
• In the event destructor, the backend event is returned to the event pool.

This should allow us to elide most calls to event creation functions.

CC @al42and @pszi1ard