[BUGFIX] Improve compile/use of nvmlDeviceGetComputeRunningProcesses()

[guanxingithub]

Description

This PR is filed to solve the issue reported in #20467 as well as #20863. For fixing this issue, an initial solution was provided in #20499 by @khaotik. This PR, with a fix from @DickJC123, adds more compilation robustness across drivers and both settings of the NVML_NO_UNVERSIONED_FUNC_DEFS cmake flag.

In this PR, we would like to merge these two solutions and find a general solution that avoids compilation errors no matter which signature of the nvmlDeviceGetComputeRunningProcesses() function is enabled in the code.

Checklist

Essentials

Dick Carter finished coding on this PR
All changes are fully tested by Guanxin Li
Code is well-documented

Changes

The change reduce the possibly versioned variant of nvmlProcessInfo_t* expected
as the 3rd arg of nvmlDeviceGetComputeRunningProcesses().

Comments

This change is a backward incompatible change, without this change, mxnet can't be compiled in cuda11 450.x driver

[mxnet-bot]

Hey @guanxingithub , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [clang, miscellaneous, unix-cpu, windows-gpu, unix-gpu, windows-cpu, edge, website, sanity, centos-gpu, centos-cpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[ptrendx]

Moving my approval from #20866

[DickJC123]

Based on the signature of the nvmlDeviceGetComputeRunningProcesses() function (as dictated by the driver include file and the NVML_NO_UNVERSIONED_FUNC_DEFS cmake flag), we want the local variable infos vector to have an element type compatible with the 3rd arg of that function. This PR uses template meta-programming to deduce that. LGTM.