cl_khr_command_buffer

[rjodinchr]

@kpet draft implementation followed by:

• Compile with CL_ENABLE_BETA_EXTENSIONS
• Complete implementation for new cl*CommandBufferKHR functions
• Complete implementation for new DeviceInfo entries
• Commands can be reused:
  • Reset event
  • Remove command from event
  • Move QueryPool inside event instead of command
• Add optional support to fill image on device (used by default for
  command buffer)
• Refcount command for easier lifetime management
• cvk_command_batch holds both commands and events as getting event
  through command cannot be done once the command has been executed.
• Add 'clone' method for all commands for queue migration

[rjodinchr]

As discussed, forcing read/write flags for all images to support fill on device is not great as it can have a significant impact on performance for some workloads.
On top of that the implementation is either breaking test_api negative_set_read_write_image_arg or test_cl_khr_command_buffer copy_image. One requires to return an error for image not matching the access qualifier, while the other has a read_only image that is getting fill on device, thus it gets passed to the kernel using a write_only qualifier.

[rjodinchr]

I have updated that PR with a new implementation that does not require all images to use read/write flags. It creates a second image (with write_only flag) using the same VkDeviceMemory at runtime when it detects a fillimage on a read_only image.

[rjodinchr]

Marking as draft while the fill image on device PR is under review.