updated QDQ config for SqueezeNet

[xieofxie]

Summary

Closes properties_not_found gaps surfaced by a QNN/NPU coverage scan of SqueezeNet. All gaps collapsed to one pattern — dangling QDQ: nodes whose inputs or outputs are not wrapped by a DQ/Q node in the real model, producing filters like QDQ_Y=None or QDQ_A=None that had no matching row in the EP's support table.

Post-scan on SqueezeNet_opt.onnx → 0 gaps.

Generator changes

File	Change	Gap addressed
binary_input_generator.py	BinaryInputGenerator.get_qdq_config() — add support_non_qdq=True on A and B	Mul with QDQ_A=None, QDQ_B=None (raw-float inputs feeding a quantized Mul). Inherited by Add/Sub/Mul/Div/Pow/And/Or/Bitwise*/PRelu/BitShift/Mod. Comparison and Where override and are unaffected.
flatten_input_generator.py	Add "output" key with support_activation=True, support_non_qdq=True	Flatten with QDQ_output=None
global_pooling_input_generator.py	Add "Y" key with the same two flags	GlobalAveragePool with QDQ_Y=None
conv_input_generator.py	Add SupportedONNXType.INT8 to B's qdq_types	Conv with QDQ_B=INT8 (INT8-quantized bias observed in P1 models). Marked with a TODO — revisit once EP-level INT8 bias support is confirmed/rejected.

Test count updates

tests/unit/analyze/core/test_qdq.py::test_qdq_total_count — re-derived with per-bucket decompositions so the math is explicit:

• Add → binary_input_shapes * 80 = 3440, broken down by (A, B) should_qdq combo: 28 (Q,Q) + 20 (Q,raw) + 20 (raw,Q) + 12 (raw,raw).
• Conv → 1536 * 5 = 7680, five bias states (no bias / INT32 / INT8 / non-QDQ constant / non-QDQ non-constant) each yielding 1536 models.
• Flatten → 28 * 4 * 2 = 224 (output Q-wrapped or not).
• GlobalAveragePool → 3 * 4 * 2 = 24.

Full tests/unit/analyze/ suite green (1341 passed, pre-existing skips unchanged).