Support static linking of libcudnn

[take-cheeze]

Since most FindCUDNN.cmake's static feature seems to be broken from cudnn v8 I've like to fix it.
libcudnn_static.a isn't included at least from v9 and there is more tricks needed to link static cudnn documented in:
https://docs.nvidia.com/deeplearning/cudnn/installation/latest/build-run-cudnn.html#running-a-cudnn-dependent-program

In this PR, I'll add 2 hint flags:

• CUDNN_STATIC to enabled static linking feature
• CUDNN_SKIP_PRECOMPILED_LINK to disable linking to cudnn_engines_precompiled which sometimes cause overflowing of 2GiB relocation limit of libcudnn's 32bit binary which is reported in https://discuss.pytorch.org/t/libtorch-cxx11-abi-static-library/211309

I've tried

cmake -GNinja -Bbuild -S. -DCUDNN_STATIC_LINK=ON -DCUDNN_SKIP_PRECOMPILED_LINK=ON
cmake --build build
./build/bin/samples

working and result of ldd seems to be minimum as I expect:

$ ldd ./build/bin/samples 
        linux-vdso.so.1 (0x0000789a47e99000)
        libcudart.so.12 => /usr/local/cuda-12.8/targets/x86_64-linux/lib/libcudart.so.12 (0x00007899f2c00000)
        libcuda.so.1 => /lib/x86_64-linux-gnu/libcuda.so.1 (0x00007899ece00000)
        libstdc++.so.6 => /lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007899eca00000)
        libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007899f2f17000)
        libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007899f2ee9000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007899ec600000)
        /lib64/ld-linux-x86-64.so.2 (0x0000789a47e9b000)
        libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x0000789a47e67000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x0000789a47e62000)
        librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007899f2ee4000)