Use SnoopPrecompile

[mcabbott]

This aims to reduce startup time using https://github.com/timholy/SnoopCompile.jl, as described in this thread.

So far uses existing warm-up cases, but perhaps it should target many of the runtime functions too.

[mcabbott]

Moving to StaticArraysCore might also be worthwhile, since this is the bulk of this time:

julia> @time_imports using TensorCast
     13.3 ms  MacroTools
      2.7 ms  StaticArraysCore
    573.2 ms  StaticArrays
      0.9 ms  TupleTools
     33.2 ms  Strided 37.04% compilation time
      0.9 ms  DelimitedFiles
      0.3 ms  Compat
     83.3 ms  ChainRulesCore
      0.3 ms  Adapt
      2.9 ms  GPUArraysCore
      0.2 ms  Requires
     23.6 ms  TransmuteDims 82.86% compilation time (10% recompilation)
      1.4 ms  LazyStack
    132.9 ms  TensorCast