Apply constant folding to Q/DQ patterns on constants

[pluralia]

The update is required to remove Slice and Transpose ops if MM is identified as act x weight instead of act x act.

Adds three new patthers to src/Dialect/ONNX/Transforms/ConstProp.cpp, all gated behind the existing enableQDQ flag and registered alongside the current RemoveQDQForConst. They target Q/DQ chains rooted at an ONNXConstantOp that the existing patterns leave behind:

• BypassShapeOpThroughDQ<ONNXOp> — swaps Const → DQ → ShapeOp into Const → ShapeOp → DQ for Slice/Transpose/Reshape/Squeeze/Unsqueeze so the existing *OfConst folders can collapse the shape op on the integer constant.

• DropIdempotentQDQOnConst — removes a Const → DQ(s,z) → Q(s,z) round-trip on an integer constant when the DQ and Q share the same scale, zero-point, and storage dtype

• FoldRequantizeOnConst — materializes a new integer constant for a genuine Const → DQ(s1,z1) → Q(s2,z2,intT2) requantize by recomputing each element with round-to-nearest-even and saturation

[Copilot]

Pull request overview

This PR extends ONNX constant propagation to further fold/remove Q/DQ patterns rooted at constants (under the existing enableQDQ gate), helping eliminate leftover Slice/Transpose/etc. that can block later optimizations.

Changes:

• Adds BypassShapeOpThroughDQ<...> to commute shape-only ops ahead of DequantizeLinear for per-tensor quantization on constants.
• Adds DropIdempotentQDQOnConst to remove no-op Const -> DQ -> Q round-trips on integer constants when parameters are identical.
• Adds FoldRequantizeOnConst to materialize a new integer constant for genuine Const -> DQ -> Q requantization.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[p-lanza]

please add unit tests

[guwacAMD]

DMAC FE Regression Summary

• 151 models submitted, 145 pass / 6 fail on both sides — pass/fail set identical.
• No new failures, no recovered passes.
• 6 failures pre-exist on nightly (unrelated to the patch):
  • exit 137 (OOM): fara-qwen-vl-1, fara-qwen-vl-512, phi-v-next-7b-psu0, phi-v-next-7b-psu1
  • exit 134 (abort): mako-encoder-fp16, psr

Per-model frontend diff

• 143 / 145 identical between CHK and REF.
• 2 / 145 show real diffs: psx0-qdq (target) and microsoft-infoxlm-large (side effect).

psx0-qdq — net op-type deltas (REF → CHK)

op_type	REF	CHK	Δ
MatMul_qdq_actxact_uint16xuint16xuint16	120	0	−120
MatMul_qdq_uint16xuint8xuint16	0	120	+120
MatMul_qdq_bias_uint16xuint8xuint16	0	2	+2
Slice_qdq_uint8xbfloat16	120	0	−120
Transpose_qdq_bfloat16xuint16	120	0	−120
Transpose_qdq_uint16xuint16	3	1	−2
Transpose_qdq_uint8xuint8	4	0	−4
Dequant_uint16xbfloat16	4	0	−4
Dequant_int16xbfloat16	2	0	−2
Quant_bfloat16xuint16	2	0	−2

120 actxact uint16×uint16 MatMuls reclassified to AIE-runnable actxweight uint16×uint8; 120 paired Slice + 120 Transpose ops on weights collapsed; 2 1×1 Convs (AIESW-34318) now fused as MatMul_qdq_bias_uint16×uint8.

microsoft-infoxlm-large — side effect

• 1 Conversion_uint8xuint16 removed (lm_head.dense.bias_DequantizeLinear)
• 1 LayerNormalization_qdq weight retyped from uint16 → uint8 (B_dtype, wgt1_bytes 2→1)
• Net: −1 op, small improvement