[build] Don't build for Tegra sm_XX versions on x86/ppc and vice versa; allow --cuda-arch overrides to have multiple versions (#3171)

was only apparent using large matrices

clang is unhappy with '-rdynamic' in compile-only step, and the
switch is really unnecessary.

Also, the default location for MKL 64-bit libraries is intel64/.
The em64t/ was explained already obsolete by an Intel rep in 2010:
https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/285973

[src] Fix to older online decoding code in online/ (OnlineFeInput; was broken by commit cc2469e8). (#3025)