Update triton-shared submodule and remove deprecated TritonToLinalg pass

Update triton_shared submodule from 1af6a5f to 209a064 (latest
facebookincubator/triton-shared main), which removes the deprecated
monolithic TritonToLinalg pass. Regenerate triton_shared.patch to
remove redundant TritonToLinalg deletion hunks now handled upstream.

The amd_triton_npu backend exclusively uses --triton-to-linalg-experimental
(TritonArithToLinalg + TritonToLinalgExperimental), so this removal has
no functional impact. Also update README triton-shared link to point to
the facebookincubator fork.

Tested: full build + matmul on NPU hardware (AIE2) — identical results
to unmodified main.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: erwei-xilinx

README.md

@@ -6,7 +6,7 @@ Triton-XDNA provides an end-to-end compilation flow that lowers standard Triton
 
 ### How it works
 
-Triton kernels are first lowered to compact Linalg compute graphs via [triton-shared](https://github.com/microsoft/triton-shared), then tiled and mapped onto parallel NPU cores using the MLIR Transform dialect, and finally compiled through [MLIR-AIR](https://github.com/Xilinx/mlir-air) and [MLIR-AIE](https://github.com/Xilinx/mlir-aie) to produce device binaries.
+Triton kernels are first lowered to compact Linalg compute graphs via [triton-shared](https://github.com/facebookincubator/triton-shared), then tiled and mapped onto parallel NPU cores using the MLIR Transform dialect, and finally compiled through [MLIR-AIR](https://github.com/Xilinx/mlir-air) and [MLIR-AIE](https://github.com/Xilinx/mlir-aie) to produce device binaries.
 
 ```
 Triton kernel (@triton.jit)