The ARM core on Kindle Fire has 128-bit registers in addition to the usual 32-bit registers. This could give a 2-4 times speedup, depending on the exact details of the core used.
Unfortunately not. I have already written a version of the scrypt core that uses NEON (it will be released in the upcoming cpuminer 2.2.3), and the speedup is only about 40% in the best case. There are various technical reasons for this, most importantly the lack of an instruction for bitwise rotation and the fact that current NEON coprocessors can only execute one instruction per cycle, with rare exceptions.
The basic ARM instruction set, on the other hand, is pretty optimal for implementing this kind of hashing functions (thanks to the barrel shifter), but of course it cannot compete with the SIMD implementations of performance-oriented architectures.
An example: a Samsung Galaxy S II (dual-core Cortex-A9 at 1200 MHz) can do about 1.64 khash/s with the current cpuminer code (hand-written ARM assembly), and about 2.3 khash/s with the to-be-released NEON implementation. The speedup is even more modest on Cortex-A8 processors.
this is important stuff, and I guess shows how an arm based system designed to mine litecoin might perform. at 125w, about 166 arm cortex a9 cpus could thus mine only approximately 210khs if I'm doing the math right (cortex cpu ~ 750mW when running neon code fully loaded).
which in 3 days will be up. I'm currently on vacation with the family... be back on Thursday or Friday. 
