Remove pystencils.GPU_DEVICE

This function reports the sizes of L1/L2/L3 caches and free and total RAM, so why not also have it report the memory size of each GPU instead of just the selected GPU. It is not used anywhere, so the slight API change has no consequences.

Related to pystencils!336 (merged)

test_gpu_block_size_limiting.py::test_gpu_block_size_limiting actually failed on master since pystencils!335 (merged), but that is due to an error in the test (it was expecting the input argument to be modified and ignoring the return value of limit_block_size_by_register_restriction).

lbmpy/max_domain_size_info.py

@@ -26,7 +26,6 @@ Examples:
 import warnings
 
 import numpy as np
-import pystencils
 
 # Optional packages cpuinfo, cupy and psutil for hardware queries
 try:
@@ -36,19 +35,8 @@ except ImportError:
 
 try:
     import cupy
-    device = cupy.cuda.Device(pystencils.GPU_DEVICE)
 except ImportError:
     cupy = None
-    device = None
-
-if cupy:
-    try:
-        device = cupy.cuda.Device(pystencils.GPU_DEVICE)
-    except AttributeError:
-        warnings.warn("You are using an old pystencils version. Consider updating it")
-        device = cupy.cuda.Device(0)
-else:
-    device = None
 
 try:
     from psutil import virtual_memory
@@ -125,9 +113,11 @@ def memory_sizes_of_current_machine():
         if 'l3_cache_size' in cpu_info:
             result['L3'] = cpu_info['l3_cache_size']
 
-    if device:
-        size = device.mem_info[1] / (1024 * 1024)
-        result['GPU'] = "{0:.0f} MB".format(size)
+    if cupy:
+        for i in range(cupy.cuda.runtime.getDeviceCount()):
+            device = cupy.cuda.Device(i)
+            size = device.mem_info[1] / (1024 * 1024)
+            result[f'GPU{i}'] = "{0:.0f} MB".format(size)
 
     if virtual_memory:
         mem = virtual_memory()