fix(non_uniform_work_group): respect per-dimension work item size lim…

[Lurie97]

…its when selecting localSize

Problem:
Test cases selected localSize based only on CL_KERNEL_WORK_GROUP_SIZE, without considering CL_DEVICE_MAX_WORK_ITEM_SIZES per-dimension limits.

This caused issues because:

1. Constructor rounds globalSize based on the original localSize
2. prepareDevice() later trims _enqueuedLocalSize to device limits
3. Result: globalSize was rounded with wrong localSize, causing mismatch

For example, with maxWorkItemSizes=[256,256,64] and localSize={512,1}:

• globalSize rounded assuming localSize[0]=512
• prepareDevice() limits localSize[0] to 256
• globalSize and localSize no longer match

Fix:
Query CL_DEVICE_MAX_WORK_ITEM_SIZES in test cases and compute:
effectiveMax = min(maxWgSize, maxWorkItemSizes[dim])
before selecting localSize for each dimension.

This ensures localSize is valid before rounding, so globalSize and localSize remain consistent throughout the test.

Affected files:

• test_advanced_2d.cpp
• test_advanced_3d.cpp
• test_advanced_other.cpp
• test_basic.cpp

[CLAassistant]

All committers have signed the CLA.

[rjodinchr]

I wonder if #2598 is fixing the same issue or not.

[Lurie97]

Yes, PR #2598 fixes the same issue. That said, I believe both changes can coexist. My patch addresses the root cause at test generation time, while the prepareDevice() change provides a runtime safeguard. Together, they improve robustness by both preventing and handling the issue.