Fix Triton build using icx on Windows (#5224)

# Fix mirror bugs in compilation

1. ```partitions.getPartition((int)0)```
`getPartition` function has two type parameters(`unsigned` and
`Operation *`)
  There are no candidates for `int`.

2. ```constexpr int64_t SRC_MASK = (1ULL << (SrcBits - 1)) - 1;```
On Windows/MSVC, long is 32-bit (LLP64). When `SrcBits == 32`, this
becomes `1L << 31`. Left-shifting a signed 32-bit long into the sign bit
is **undefined behavior**. MSVC treats this as non–constant at compile
time, causing: `constexpr variable must be initialized by a constant
expression`
Use an unsigned 64-bit base for bit shifts (and make masks unsigned), so
the expression is well-defined and remains a constant expression across
compilers

Author: liaojianjin

lib/Dialect/TritonGPU/Transforms/WarpSpecialization/PartitionLoops.cpp

@@ -410,7 +410,7 @@ LogicalResult triton::gpu::partitionLoop(scf::ForOp loop) {
     return success();
 
   auto numPartitions = partitions.getNumPartitions();
-  auto defaultPartition = partitions.getPartition((int)0);
+  auto defaultPartition = partitions.getPartition((unsigned int)0);
   auto loopVarCategories = classifyLoopVars(loop, defaultPartition, partitions);
   auto [loopVarIndices, newResultIndices] =
       getLoopVarIndicesToKeep(loop, defaultPartition, loopVarCategories);

third_party/intel/lib/TritonIntelGPUToLLVM/ElementwiseOpToLLVM.cpp

@@ -458,7 +458,7 @@ static SmallVector<Value> Fp_to_Fp8_RTNE(Location loc,
                 (std::is_same_v<SrcTy, BFloat16Type>));
   static_assert((std::is_same_v<DstTy, Float8E4M3Type>) ||
                 (std::is_same_v<DstTy, Float8E5M2Type>));
-  constexpr int64_t SRC_MASK = (1L << (SrcBits - 1)) - 1;
+  constexpr int64_t SRC_MASK = (1ULL << (SrcBits - 1)) - 1;
   constexpr int64_t SRC_MMASK = (1L << SrcMBits) - 1;
   constexpr int64_t DST_NAN = 0x7F;
   constexpr int64_t DST_MAX =