[stdlib][GPU] Fix invalid constant addr space

On Nvidia, address space value of 2 in the IR is reserved.  So, it's
invalid for us to represent the constant address space with the value of
`2`.

Good news: nobody is setting this value currently.  So, just remove it.
While here, rename the existing `PARAM` address space name to be called
"CONSTANT" instead, which is more consistent with the
[docs](https://docs.nvidia.com/cuda/nvvm-ir-spec/#address-space).  Note
that AMD calls address space 4 "constant" as well.

MODULAR_ORIG_COMMIT_REV_ID: 35bbc8ad2b82efdddf6fa4cf537132af2ca2d3af

stdlib/src/memory/memory.mojo

@@ -414,9 +414,12 @@ fn stack_allocation[
 
     @parameter
     if is_gpu():
-        # On NVGPU, SHARED and PARAM address spaces lower to global memory.
+        # On NVGPU, SHARED and CONSTANT address spaces lower to global memory.
         @parameter
-        if address_space in (_GPUAddressSpace.SHARED, _GPUAddressSpace.PARAM):
+        if address_space in (
+            _GPUAddressSpace.SHARED,
+            _GPUAddressSpace.CONSTANT,
+        ):
             alias global_name = name.value() if name else "_global_alloc"
             return __mlir_op.`pop.global_alloc`[
                 name = global_name.value,

stdlib/src/memory/pointer.mojo

@@ -38,12 +38,10 @@ struct _GPUAddressSpace(EqualityComparable):
     """Generic address space."""
     alias GLOBAL = AddressSpace(1)
     """Global address space."""
-    alias CONSTANT = AddressSpace(2) if is_nvidia_gpu() else AddressSpace(4)
-    """Constant address space."""
     alias SHARED = AddressSpace(3)
     """Shared address space."""
-    alias PARAM = AddressSpace(4)
-    """Param address space."""
+    alias CONSTANT = AddressSpace(4)
+    """Constant address space."""
     alias LOCAL = AddressSpace(5) if is_nvidia_gpu() else AddressSpace(3)
     """Local address space."""