accumulate on Metal sometimes fails due to weaker @synchronize guarantees than on other platforms

[anicusan]

Possible fixes:

• From @maleadt on Slack: "the default KA.jl @synchronize for Metal.jl calls threadgroup_barrier with flag=Metal.MemoryFlagThreadGroup, i.e., it only orders memory operations to threagroup memory. If it's device memory, you could try MemoryFlagDevice"
• Implement algorithm without decoupled opportunistic lookback.

[anicusan]

Leaving here as a note: using Metal.MemoryFlagDevice did not help (though it probably makes it slower); the problem is that Metal doesn't flush caches of device arrays after synchronize statements. It seems opportunistic lookback just cannot be implemented on Macs (see post and issue); turns out they had to change the WebGPU standard because of it. Isn't parallelism fun when your algorithm has a 0.38% chance of returning cobblers on just one out of 5 backends?