fix(emitc): avoid invalid static-to-dynamic GlobalTensor cast for MGATHER/MSCATTER

[FangRui0]

Summary

Fixes #1165 — a v0.55 regression where PTO-to-EmitC lowering emits C++ that fails to compile for MGATHER / MSCATTER.

Root cause. TensorViewType / PartitionTensorViewType carry only shape + element type in the MLIR type — never strides. So PTOToEmitCTypeConverter can only map them to a GlobalTensor opaque type with a fully-dynamic Stride<-1,-1,-1,-1,-1> template. The static partition_view pattern, which does have access to the defining op, instead materializes a GlobalTensor with concrete static strides (e.g. Stride<2048,2048,2048,32,1>). Dialect conversion bridges the two mismatched GlobalTensor types with an unrealized_conversion_cast, and the cleanup lowered it to an emitc.cast — an invalid C-style cast between two GlobalTensor instantiations that have no converting constructor:

GlobalTensor<...,Stride<-1,-1,-1,-1,-1>,...> v19 =
    (GlobalTensor<...,Stride<-1,-1,-1,-1,-1>,...>)v18;   // no matching conversion
MGATHER<pto::Coalesce::Row>(dst, v19, idx);

This surfaced in v0.55 once MGATHER/MSCATTER became tensor-view-native and the partition-view GlobalTensor started flowing directly into them (commits 61ed749, 83901f6, 8d1d8b5).

Fix

In the unrealized-cast cleanup, recognize a GlobalTensor → GlobalTensor bridge that is identical except for static-vs-dynamic Shape/Stride template params (areRefinableGlobalTensorTypes) and forward the value instead of emitting a cast. The static-stride GlobalTensor then flows straight into the templated MGATHER/MSCATTER call, which accepts any instantiation. A C-style cast between differing GlobalTensor template instantiations is never valid C++, so this cleanup path should never produce one.

Emitted C++ after the fix (no intermediate cast, static strides preserved):

GlobalTensor<float, pto::Shape<1,1,1,64,32>, pto::Stride<2048,2048,2048,32,1>, pto::Layout::ND> v18 = ...;
MGATHER<pto::Coalesce::Row>(v14, v18, v9);   // consumes v18 directly

Validation (remote A3, LLVM21 toolchain)

• Full build 443/443.
• llvm-lit test/lit/pto/mgather_mscatter_base_a3_emitc.pto — PASS, including the new NOCAST guard.
• Issue [Pass Bug] v0.55 emits invalid static-to-dynamic GlobalTensor cast for MGATHER/MSCATTER #1165 minimal reproducer compiled with ccec --cce-aicore-arch=dav-c220-vec -D__NPU_ARCH__=2201 -fsyntax-only — 0 errors; the original no matching conversion for C-style cast is gone and MGATHER instantiates with the static GlobalTensor type.

Test plan

• mgather_mscatter_base_a3_emitc.pto lit passes (adds a second FileCheck pass NOCAST-NOT: (GlobalTensor< guarding against re-emission of the C-style cast)
• Generated C++ for the [Pass Bug] v0.55 emits invalid static-to-dynamic GlobalTensor cast for MGATHER/MSCATTER #1165 reproducer compiles cleanly (ccec syntax/type-check, 0 errors)
• Full check-pto regression on CI