Win: Use DXGI for accurate WDDM VRAM memory budget

[dhiltgen]

WDDM implements VRAM paging, which can result in a performance cliff at the limit. This wires up DXGI to expose more accurate VRAM budget.

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: co-developed with a coding agent

[dhiltgen]

This slightly contrived example runs mlx-lm and evicts old caches under memory pressure.
On a 5090 using Llama-3.2-3B-Instruct-4bit, you can toggle by adding/removing --retain-all

--conversations 31 --prompt-tokens 8192 --generation-tokens 4 --prefill-step-size 512 --reserve-gib 2 --retain-all

mlx-lm-wddm-cache.py

[zcbenz]

The allocator of MLX is currently not aware of multi-device and just assumes all devices share same memory size and limits. So this function should return per-device limit instead of accumulation.

Suggested change

	size_t windows_memory_limit(
	size_t memory_limit,
	const std::vector<cudaMemPool_t>& pools) {
	static WddmMemoryBudget budget(pools.size());
	return budget.get_memory_limit(memory_limit, pools);
	}
	size_t get_windows_memory_limit(
	size_t hard_memory_limit,
	int device,
	cudaMemPool_t pool) {
	static WddmMemoryBudget budget(device_count());
	return budget.get_memory_limit(hard_memory_limit, device, pool);
	}