Fix state transitions for re-running requests

[rmccorm4]

Fixes L0_simple_lib--base, and allows requests to be re-used for multiple inferences.

Adds a sanity check that pending count isn't incremented multiple times without getting decremented in some edge case where a request is PENDING, fails to get to the PENDING state, and ends up in PENDING state again (re-used).

General logic:

1. When a request is created, it is in INITIALIZED state
2. When a request is run, it is set to PENDING state
3. If a request makes it to the backend, it should be in EXECUTING state
   • At this point, the backend takes ownership of the request, and is responsible to call RequestRelease() on it, moving request into RELEASED state
4. If any error occurs before getting to backend, it should hit InferenceRequest::RespondIfError and be set to RELEASED state.
5. If somehow a request didn't make it to backend (EXECUTING state), but also didn't get RELEASED, then it should still be in PENDING state. If the request is to be re-used for another inference, it should call PrepareForInference() which should set state back to INITIALIZED at the start of the cycle.

If a request were to be transitioned from RELEASED to PENDING/EXECUTING directly, this should indicate a logical error as the request should only execute via InferAsync entrypoint, which calls PrepareForInference ([re]sets to INITIALIZED state) then InferenceRequest::Run (sets to PENDING state).