Remove the per-op completion handler in gpu::eval

[kernelpool]

Proposed changes

gpu::eval registers an empty addCompletedHandler and builds an unordered_set of shared_ptrs for every operation, in order to keep that op's inputs alive until the command buffer completes. CommandEncoder::commit() already installs one handler per command buffer, so the per-op handlers are redundant and the buffers can go there instead. On an M3 Ultra that takes mlx-community/Qwen3.5-0.8B-4bit decode from 280 to 367 tok/s. Bandwidth-bound models are unchanged.

Before

mlx_lm.benchmark --model mlx-community/Qwen3.5-0.8B-4bit -p 2048 -g 128 -n 3
Running warmup..
Timing with prompt_tokens=2048, generation_tokens=128, batch_size=1.
Trial 1:  prompt_tps=13108.539, generation_tps=284.374, peak_memory=1.886, total_time=0.744
Trial 2:  prompt_tps=13404.757, generation_tps=278.961, peak_memory=1.886, total_time=0.749
Trial 3:  prompt_tps=13181.042, generation_tps=277.251, peak_memory=1.887, total_time=0.753
Averages: prompt_tps=13231.446, generation_tps=280.195, peak_memory=1.886

After

mlx_lm.benchmark --model mlx-community/Qwen3.5-0.8B-4bit -p 2048 -g 128 -n 3
Running warmup..
Timing with prompt_tokens=2048, generation_tokens=128, batch_size=1.
Trial 1:  prompt_tps=13288.070, generation_tps=365.077, peak_memory=1.886, total_time=0.637
Trial 2:  prompt_tps=13271.215, generation_tps=361.761, peak_memory=1.886, total_time=0.650
Trial 3:  prompt_tps=13402.103, generation_tps=376.355, peak_memory=1.887, total_time=0.629
Averages: prompt_tps=13320.463, generation_tps=367.731, peak_memory=1.886

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[zcbenz]

I think you can just store buffers in the set.

Suggested change

	std::vector<std::shared_ptr<array::Data>> retained_buffers_;
	std::unordered_set<const array::Data*> retained_ptrs_;
	std::unordered_set<std::shared_ptr<array::Data>> retained_buffers_;

[zcbenz]

The downside of this approach is that buffers are now retained until the command encoder is committed, which I think could increase peak memory a lot? Can you check the memory usage in a few popular models?

I support reducing the completion handler to have lower overhead but we probably need to add some heuristics.