Geekbench tests daily tasks like AES, PDF, HTML5. If the score is higher, then daily tasks are faster too.
IMHO it is reliable across different Arch/CPU/OS.
If anything, I feel like many tests are based on x86 special instructions like AES, which does not directly translate to other tasks. But for most of the tasks I think it wouldn't be possible that the processor are only better at those and performance of most program should correlate with geekbench score.
ARM has AES instruction extensions, and IIRC iPhones have had them for some time.
Generally, yes. The benchmarks are calculating/computing the same things on both architectures, and the various sub-benchmarks are based on real-world non-trivial computational problems rather than microbenchmarks that are easily manipulated by instruction set differences.
Unless they are written in optimised Asm I would not compare across architectures, because there is considerable leeway in compilers and my experience has shown that a good human can often beat a compiler solidly --- at least on x86.