FFT: 100x speedup by cutting the BS

[AntonOresten]

The tiles currently have a batch dimension with size equivalent to the array itself. This means that only the first block is doing useful work, and the kernel slows down massively with problem size as things start to spill.

In this PR I interpret the "BS" kernel argument as a minibatch, and adjust the grid size accordingly, as well as cleaning up some integer arguments that will constant-fold anyway.

On the default problem size for benchmarks, it nows runs ~5x faster than the cuFFT (without a plan?) instead of ~25x slower. Timed on a DGX Spark. It isn't consistently faster than cuFFT though, especially for larger factors.

For the benchmarks, the problem size should definitely be updated as we're in the ~10 microsecond range with this fix, and the relative performance over cutile-python (currently reported as ~equal) may or may not hold up.

Ideally we'd drop the BS entirely, as I'm seeing BS=1 is the fastest. But it might still offer slight performance benefits at small problem sizes, and makes the kernel more expressive.

[maleadt]

Great improvements, thanks!

Before:

========================================================================
  fft
========================================================================
Implementation      Min (ms)    Mean (ms)   Throughput
------------------------------------------------------------------------
cuFFT               0.065       0.067       65 μs
cuTile              0.543       0.55        543 μs
------------------------------------------------------------------------

After:

========================================================================
  fft
========================================================================
Implementation      Min (ms)    Mean (ms)   Throughput
------------------------------------------------------------------------
cuTile              0.011       0.012       11 μs
cuFFT               0.062       0.066       62 μs
------------------------------------------------------------------------

With plan outside of the cuFFT benchmark:

========================================================================
  fft
========================================================================
Implementation      Min (ms)    Mean (ms)   Throughput
------------------------------------------------------------------------
cuFFT               0.004       0.005       4 μs
cuTile              0.011       0.012       11 μs
------------------------------------------------------------------------

CUDA.jl's HandleCache was supposed to chache plans already, but it's GC-driven so not entirely reliable. So I pushed a commit moving the planning out.