DO NOT MERGE: New DaCe GPU Codegen

pyproject.toml

@@ -100,7 +100,7 @@ dependencies = [
   'click>=8.0.0',
   'cmake>=3.22',
   'cytoolz>=1.0.1',
-  'dace>=2.0.0a3',
+  'dace==2.3.7',
   'deepdiff>=8.1.0',
   'devtools>=0.6',
   'factory-boy>=3.3.3',
@@ -476,6 +476,9 @@ url = 'https://gridtools.github.io/pypi/'
 # dace = {index = "gridtools"}
 [tool.uv.sources]
 atlas4py = {index = "test.pypi"}
+dace = [
+  {git = "https://github.com/philip-paul-mueller/dace", branch = "phimuell__new-gpu-codegen-dev"}
+]
 
 # -- versioningit --
 [tool.versioningit]

src/gt4py/next/program_processors/runners/dace/workflow/common.py

@@ -109,6 +109,9 @@ def set_dace_config(
     #  This setting allows to throw an exception if any implicit Copy-Map slips thorugh.
     dace.Config.set("compiler.cuda.allow_implicit_memlet_to_map", value=False)
 
+    # Use the new GPU code generator
+    dace.Config.set("compiler.cuda.implementation", value="experimental")
+
     # In some stencils, for example `apply_diffusion_to_w`, the cuda codegen messes
     #  up with the cuda streams, i.e. it allocates N streams but uses N+1. The first
     #  idea was to use just one stream. However, even in that case the generator