No longer silently hide errors in Metal completion handlers (alternative approach)

[shoaibkamil]

As discussed in the dev meeting, here is an alternative approach for fixing #7780. Now, the runtime saves an error during the callback from a failed command buffer, and reports the error back at the next opportunity for reporting. Note that this means that a pipeline may get an error that should be attributed to the previous pipeline, but additional error information can be added (e.g. prepending the error message with the name of the kernel that caused it) if wanted.

[shoaibkamil]

Sorry @steven-johnson, I brain-farted and hadn't pushed the branch I intended to push when I opened this. Should be correct now.

[slomp]

@shoaibkamil I recommend pre-pending the error message with something that reminds the user that this error has happened unexpectedly during the execution of a previous pipeline.

[steven-johnson]

C:\build_bot\worker\halide-testbranch-main-llvm19-x86-64-windows-cmake\halide-source\test\correctness\gpu_metal_completion_handler_error_check.cpp(2): fatal error C1083: Cannot open include file: 'unistd.h': No such file or directory

[shoaibkamil]

@steven-johnson I've substantially revamped the API here to enable users to pass a user context. PTAL

[steven-johnson]

LGTM pending green buildbots

[shoaibkamil]

Investigating crash, and will add tests to makefile.

[shoaibkamil]

All green finally, modulor vulkan errors.