thunk,Sch: Reduce invalidations and TTFX

[jpsamaroo]

Closes #333

[jpsamaroo]

@DrChainsaw I now get TTFX (with 6 workers):

julia> testfun(2000);
  2.424907 seconds (1.48 M allocations: 75.603 MiB, 0.66% gc time)
  0.061711 seconds (117.94 k allocations: 4.473 MiB)

We still see increasing TTFX as the size of the vector being splatted increases, but that will be hard to fix; and even if we do, the scheduler still does work that is approximately O(N), so that will impede scaling.

[DrChainsaw]

Awesome! I see a sizeable improvement as well. In particular the recompilation when number of thunks change is gone now (I guess thanks to a couple of Tuple -> Vectors).

# Very first call
julia> testfun(2000);
 23.585389 seconds (21.31 M allocations: 1.153 GiB, 2.80% gc time, 47.90% compilation time)
  2.028270 seconds (1.05 M allocations: 55.230 MiB, 1.06% gc time, 43.95% compilation time)

julia> testfun(2000);
  2.978357 seconds (1.83 M allocations: 122.209 MiB, 1.81% gc time)
  0.284688 seconds (128.13 k allocations: 4.786 MiB)

julia> testfun(2001);
  2.038875 seconds (1.87 M allocations: 122.961 MiB, 1.59% gc time)
  0.218601 seconds (186.75 k allocations: 9.577 MiB, 12.06% gc time)