Speed up task resolution and planning logic

[lihaoyi]

Fixes #4129 and #4026. We never really looked at performance at this stuff before, but in order to support selective.prepare __ (the default of selective.prepare) we need to improve these bottlenecks

• Introduces the ResolveCore#Cache class that holds mutable dictionaries to memo-ize various expensive parts of the task resolution logic

• Lots of micro-optimizations in various places: cache Segments#hashCode, avoid calling Segments#render for sorting or Segments#part for taking the last part, etc.

Brings down time ./mill plan __ from ~60s to ~3s