`cuda.core.launch()` improvements

[leofang]

• Support ctypes/numpy structs
  • make sure ctypes is deprioritized
• Support converting arbitrary objects to StridedMemoryView
• Benchmarking
  • measure launch() overhead
  • reimplement type dispatcher via dict lookup instead of conditional branching

[leofang]

• Support converting arbitrary objects to StridedMemoryView

If we make this a two-step approach, where the 1st step is to simply accept passing a StridedMemoryView object to launch() as a kernel launch argument, then this would be doable once #180 (comment) is done.