Address possible memory leak during model sleep/unload

[glaziermag]

Tentatively addressing issue #545.

It appears that the VRAM memory leak during /v1/sleep and model unload operations might be related to asynchronous CUDA context bindings executing after the OS thread has shut down or detached from the device pool. By synchronizing the engine_handler thread and explicitly pre-binding the candle_core context to the local HTTP thread before the unmapped Rust pointers evaluate their destructors, the driver seems to successfully purge the PagedAttention memory pools back to OS limits.

Still testing the broader implications of this driver context synchronization, so I'm open to feedback on whether this is the optimal approach for the architecture.

Diagnostic Traces (L4 Instance)

Before synchronization:

Sending POST to /v1/sleep...
{"model_id":"HuggingFaceTB/SmolLM-135M","status":"unloaded"}
20239 MiB

(Driver executes pointer drops synchronously but execution appears to drop unmapped, leaving memory allocated.)

After synchronization:

Sending POST to /v1/sleep...
{"model_id":"HuggingFaceTB/SmolLM-135M","status":"unloaded"}
271 MiB

(Memory returns to the intrinsic context reservation block, clearing the dynamically loaded KV cache).