analyze: resolve --ep/--device auto to a single best target

[fangyangci]

What

winml analyze previously treated --ep auto / --device auto as "fan out over every locally-available EP/device pair", which was inconsistent with build / run / perf (they resolve a single best target). This PR makes analyze's auto resolution single-target, reusing the shared sysinfo helpers (resolve_device / resolve_eps).

--ep all / --device all keep their full fan-out semantics — unchanged.

Resolution rules (per axis)

value	resolves to
auto	a single best target via shared resolve_device() / resolve_eps() — local-availability based (device priority NPU > GPU > CPU, EP priority IHV-first)
all	the full rule-data-backed set (fan-out) — unchanged
concrete (qnn, gpu, …)	itself

The cartesian product + EP_SUPPORTED_DEVICES filtering is unchanged, so mixed combos like --ep all --device auto work naturally (all EPs × the one resolved device). The old local-pairs intersection / half-auto warning block is removed — single-target auto is local by construction.

Behavior changes

command	before	after
analyze (default auto / auto)	all local pairs	1 pair (best device + its best EP)
--ep openvino (device auto)	all OpenVINO devices + warning	1 pair
--ep auto --device gpu	all GPU EPs + warning	1 pair
--ep all --device all	full matrix	unchanged

Breaking: auto now requires local hardware

Because auto resolves from local availability (like build / run), pinning a non-local EP with --device auto now exits 2 instead of warning-and-running:

$ winml analyze -m model.onnx --ep qnn      # on a machine without QNN
ERROR  Could not auto-select a device: Requested EP 'qnn' is not available on this system.

To statically analyze an EP you don't physically have, specify the device explicitly (--ep qnn --device npu) or use --device all.

Verified on real hardware

On a box with Intel OpenVINO (NPU + CPU) and NVIDIA TensorRT-RTX (GPU):

command	execution target(s)
analyze (default)	OpenVINOExecutionProvider (NPU)
--device gpu	NvTensorRTRTXExecutionProvider (GPU)
--ep openvino --device auto	OpenVINOExecutionProvider (NPU)
--ep openvino --device all	OpenVINO × NPU / GPU / CPU
--ep qnn	exit 2 (QNN not available)