Speed and memory optimization ideas

[andrewmilson]

With some of these I'm concerned about readability. I'd prefer to sacrifice a little bit of performance loss for readability. Also I'd like more benchmarks (which are rather minimal as of writing this) to evaluate the effectiveness of these changes.

Some initial ideas for optimisation:

• Change global allocator to http://jemalloc.net/
• Use uninitialised memory
• Cache FFT twiddle and scale buffers
• Branchless GPU kernels (MulPow is a prime candidate)
• Use async with GPU calls for CPU+GPU parallelism

Would be great to hear other ideas