MKL certainly should run OK on Atom with a gcc development system, and should give expected gains over generic compiled code. IPP should actually have been tested; you might ask on IPP forum. I don't know whether you can get an answer about MKL FFT performance without trying it yourself. If you can make a comparative performance test of your own on in-cache performance, you could hope for similar results with MKL.
1) MKL certainly ought to work with 64-bit linux, if such linux runs on your platform 2) 64-bit linux requires at least SSE2, no support for less; if MKL doesn't use SSE3 where applicable, that may be a reportable bug. SSE3 would be useful mainly for complex data types. 3) MKL is supposed to not use HT by default, see here http://software.intel.com/en-us/forums/showthread.php?t=67622 so you'll need to test to find your best mode. Current MKL has both OpenMP and sequential library options so you can set how many threads are used. Dual core Atom may be too new to have been tested but again it seems to be a bug if it doesn't give good 2 thread performance.
yes, IPP is doing special optimization optimization for Atom CPU. IPP's dispather chooses which of these processor-specific optimized libraries to use when your application makes a call into the IPP library.