Fix command queue creation in PROFANITY_DEBUG builds

[wock9000]

Symptom

Any build with -DPROFANITY_DEBUG dies immediately at startup:

failed to create command queue

The stock release build is unaffected and runs normally.

Root cause

Dispatcher::Device::createQueue passes &p to clCreateCommandQueueWithProperties:

#ifdef PROFANITY_DEBUG
	cl_command_queue_properties p = CL_QUEUE_PROFILING_ENABLE;
#else
	cl_command_queue_properties p = NULL;
#endif

#ifdef CL_VERSION_2_0
	const cl_command_queue ret = clCreateCommandQueueWithProperties(clContext, clDeviceId, &p, NULL);

clCreateCommandQueueWithProperties does not take a bitfield. Its third argument is a
zero-terminated cl_queue_properties[] list of {property name, value, ...} pairs.
&p is the address of a bare scalar, not such a list.

With -DPROFANITY_DEBUG, p == CL_QUEUE_PROFILING_ENABLE (0x2), so the runtime reads
the memory at &p as property name 0x2 followed by no value and no terminator. The
list is malformed, the call returns NULL, and the next line throws
std::runtime_error("failed to create command queue").

Why the release build is unaffected

In the release path p == 0, so &p points at a single zero word, which reads as a
valid empty property list. The call succeeds by accident. That is why this has gone
unnoticed: only the debug path puts a non-zero value in p, and every debug build has
been broken since OpenCL 2.0 headers became the norm.

The OpenCL 1.2 fallback (clCreateCommandQueue) is correct as written and is untouched,
since that function really does take a cl_command_queue_properties bitfield.

Fix

 #ifdef CL_VERSION_2_0
-	const cl_command_queue ret = clCreateCommandQueueWithProperties(clContext, clDeviceId, &p, NULL);
+	const cl_queue_properties props[] = { CL_QUEUE_PROPERTIES, p, 0 };
+	const cl_command_queue ret = clCreateCommandQueueWithProperties(clContext, clDeviceId, props, NULL);
 #else

Two lines, no other changes.

Verification

Linux / NVIDIA RTX 5090, Fedora 44, gcc 16.1.1, driver 595.71.05, OpenCL 3.0 (CUDA 13.2):

• Stock release build (unpatched master): compiles and runs clean at 2.089 GH/s.
• Stock -DPROFANITY_DEBUG build (unpatched master): aborts at startup with
  failed to create command queue.
• -DPROFANITY_DEBUG build with this patch: starts normally and emits the per-kernel
  profiling output the debug path is meant to produce.
• Release build with this patch: unchanged, no regression.

macOS (Apple OpenCL.framework): release build compiles clean with this patch. Note
that Apple's headers do not define CL_VERSION_2_0, so macOS takes the 1.2 fallback and
never reaches the changed line either way. The changed function was additionally
type-checked against the Khronos OpenCL 3.0 headers with and without
-DPROFANITY_DEBUG; both compile without warnings beyond the pre-existing
NULL-to-integer conversion on the line above.