[FLYDSL] Fix GDR reductions for FlyDSL 0.2.4

[IzacharyI]

Use raw MLIR vector operations so Qwen3.5 GDR decode compiles with the upgraded FlyDSL API.

Motivation

After upgrading the Qwen3.5 development branch to FlyDSL 0.2.4, GDR decode fails during cold kernel compilation:

TypeError: EnumMeta.__call__() takes from 2 to 3 positional arguments but 4 were given
FlyDSL 0.2.4 changed flydsl.expr.vector.ReductionOp into an expression enum, while the GDR kernel still invoked it as a raw MLIR operation. This prevents Qwen3.5-27B TP1 FlyDSL decode and CUDA Graph capture from starting.

Modifications

• Use the raw MLIR vector dialect for GDR reduction operations.
• Convert DSL vector values with _to_raw.
• Fix Q/K normalization and hidden-state reduction sites.

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.