Re-enable NumPy 2.0 semantics for add, sub, mul. (#4905)

triton-lang/triton#4589 mistakenly deactivated
these and reverted to the previous always-cast-to-int32 semantics.

Author: lezcano

python/triton/language/core.py

@@ -1909,26 +1909,20 @@ def where(condition, x, y, _builder=None):
 def add(x, y, sanitize_overflow: constexpr = True, _builder=None):
     x = _unwrap_if_constexpr(x)
     y = _unwrap_if_constexpr(y)
-    x = semantic.to_tensor(x, _builder)
-    y = semantic.to_tensor(y, _builder)
     return semantic.add(x, y, sanitize_overflow, _builder)
 
 
 @builtin
 def sub(x, y, sanitize_overflow: constexpr = True, _builder=None):
     x = _unwrap_if_constexpr(x)
     y = _unwrap_if_constexpr(y)
-    x = semantic.to_tensor(x, _builder)
-    y = semantic.to_tensor(y, _builder)
     return semantic.sub(x, y, sanitize_overflow, _builder)
 
 
 @builtin
 def mul(x, y, sanitize_overflow: constexpr = True, _builder=None):
     x = _unwrap_if_constexpr(x)
     y = _unwrap_if_constexpr(y)
-    x = semantic.to_tensor(x, _builder)
-    y = semantic.to_tensor(y, _builder)
     return semantic.mul(x, y, sanitize_overflow, _builder)