fix: Prevent deadlock in decoupled BLS and use-after-free in async_exec pybind callback

[pskiran1]

Fixes the following bugs in pb_stub.cc:

Deadlock in decoupled BLS with GPU tensors:
Stub::GetCUDAMemoryPoolAddress is called on the ParentToStubMQMonitor thread, which also delivers decoupled BLS responses (ProcessBLSResponseDecoupled). After notifying the parent, it unconditionally blocked waiting for an acknowledgment (waiting_on_stub). On the success path the parent may itself be blocked waiting for a decoupled response that this now-blocked thread is supposed to deliver, creating a circular wait. This caused occasional hangs when a BLS model first returned GPU tensors, triggering CUDA pool initialization.

Use-after-free in async_exec pybind callback:
The async callback lambda captures the local stub shared_ptr by reference ([&stub, ...]). Because stub is a local variable, it could be destroyed before the callback ran on the event loop, leaving a dangling reference and causing a use-after-free.

CI: triton-inference-server/server#8830