Bump up nvidia-ml-py requirement to 12.535.108

nvidia-ml-py==12.535.77 is a buggy version that breaks the struct for
process information, and should not be used (unless NVIDIA driver is
*also* buggy, 535.43, 535.54, and 535.86). The latest version
nvidia-ml-py==12.535.108 fixes the problem and is still compatible with
our supported drivers (R450+).

To ensure users who will install gpustat 1.2.0 have a correct version
of nvidia-ml-py version installed, we bump up the requirement.

See #160 and #161 for more details.

gpustat/test_gpustat.py

@@ -298,7 +298,8 @@ class NvidiaDriverMock:
 
         535.98: https://github.com/NVIDIA/nvidia-settings/commit/0cb3bef
             Rollback of 535.43.02.
-            `nvmlDeviceGetComputeRunningProcesses`: v3 -> v2 (???)
+            `nvmlDeviceGetComputeRunningProcesses`: v3 -> v2
+            the process _v3 API was removed (which is a regression).
             Removes field `usedGpuCcProtectedMemory` from `nvmlProcessInfo_st`
             (which is actually the v2 struct). See #161.
 
@@ -339,9 +340,9 @@ class NvidiaDriverMock:
 
     Relevant github issues:
         #107: nvmlDeviceGetComputeRunningProcesses_v2 added
-        #141: nvmlDeviceGetMemoryInfo (v1) broken for 510.39.01+
-        #161: Process information broken (not displayed) in R535 drivers
-        #160: OverflowError: Python int too large to convert to C long
+        #141: nvmlDeviceGetMemoryInfo (v1) broken in 510.39.01+
+        #161: Process information broken (not displayed) in 535.43-86
+        #160: OverflowError: Python int too large ... in 535.43-86
     """
     INSTANCES = []
 

setup.py

@@ -90,7 +90,7 @@ def run(self):
 
 
 install_requires = [
-    'nvidia-ml-py>=11.450.129',  # see #107, #143
+    'nvidia-ml-py>=12.535.108',  # see #107, #143, #161
     'psutil>=5.6.0',    # GH-1447
     'blessed>=1.17.1',  # GH-126
     'typing_extensions',