Add cmake flag USE_FATBIN_COMPRESSION, ON by default

[DickJC123]

As the size of libmxnet.so grows to near 2GB, with increased functionality and the addition of cuda architectures, we're running into link failures, e.g. see issue #17045

One technique that lowers lib size dramatically is 'fatbin compression', enabled by the nvcc options --fatbin-options -compress-all. This has been always a part of Makefile builds, but this PR adds it to the cmake builds. Specifically, this PR adds support to CMakeLists.txt for the cmake option -DUSE_FATBIN_COMPRESSION={ON,OFF}, with a default of ON for CUDA 11 builds and beyond. This PR proposes to leave existing cmake builds against 10.2 as they are, without fatbin compression, to avoid unnecessarily introducing unforeseen consequences to existing use cases.

Results of experiments building the 1.x branch with cuda11:

With cmake options -DMXNET_CUDA_ARCH="5.2 6.0 6.1 7.0 7.2 7.5 8.0" -DUSE_FATBIN_COMPRESSION=OFF, a cuda11 build fails with link error:

libmxnet.so: PC-relative offset overflow in PLT entry for void mxnet::op::mxnet_op::Kernel<...> ...

With the same above cmake options, but dropping arches 5.2 and 7.2, the build succeeds with a libmxnet.so size of 1.8GB.
Finally, with the same first cmake options -DMXNET_CUDA_ARCH="5.2 6.0 6.1 7.0 7.2 7.5 8.0" a cuda11 build (using fatbin compression then by default) succeeds with a libmxnet.so size of 750MB, so over a 2X decrease in size.

Both succeeding builds, one with fatbin compression and one without, ran the command:

time python -c "import mxnet as mx; x = mx.nd.array([1,], ctx=mx.gpu(0)); print((x+1).asnumpy()))"
[2.]

in the same time of 7.6 secs.

@samskalicky @anirudh2290 @ChaiBapchya @ptrendx

[mxnet-bot]

Hey @DickJC123 , 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: [website, windows-gpu, miscellaneous, edge, centos-cpu, unix-cpu, centos-gpu, sanity, unix-gpu, windows-cpu, clang]

---

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.

[leezu]

As we can change default options before the 2.0 release, it may be helpful to always enable USE_FATBIN_COMPRESSION? A couple of users run into the linking errors if gpu auto-detection fails and cmake defaults to building for all "common" gpu architectures.

[DickJC123]

I've adjusted the default to be always ON, regardless of CUDA version.

[samskalicky]

Thanks @DickJC123 for this great addition!