Add RAII destructors to CUDA driver objects

[badnikhil]

Cleanup of driver objects is fully manual (Buffer.release(), Context.detach(), Program.unload()), and Queue/Event have no cleanup path at all - a forgotten release() leaks device memory, streams, or modules.

Changes

• Buffer, Queue, Context, Program and Event free their CUDA handle when the last owner goes out of scope. Copies share a malloc'd atomic refcount via postblit (refcounting over @disable this(this) because the public API copies these structs by value — Queue.enqueue's Call struct, Context.pop/current, kernel args).
• The counter is malloc'd, not GC-allocated, so destructors running as GC finalizers never touch GC memory mid-sweep; destructors swallow driver errors (finalizer threads may have no current context).
• release()/unload()/detach() now drop one owner and only the last owner frees — releasing one copy no longer invalidates surviving copies, and double release or release-then-destructor is a no-op.
• Context.pop/current and UnifiedBuffer.asBuffer are non-owning views (null refcount, never destroy).
• Event gains a cuEventCreate constructor so it participates in RAII.

Tests

New hardware tests in tests/main.d (RTX 2050, sm_86, all passing with the existing suite): refcount survives copies; last-owner-frees with no double free; release-then-destructor safety; non-owning views; cuMemGetInfo-verified checks that a scoped 128 MiB Buffer's destructor really returns device memory and copies free exactly once; 1000-iteration Queue+Event create/destroy churn.

Note: the memory-verification tests use cuMemGetInfo, added to bindbc-cuda