fix type inference regression in Ryu

[KristofferC]

In #37163, a performance regression snuck in, likely due to a misunderstanding in that this array is not only used during compilation time but also at runtime. Making it Any is therefore detrimental for performance. This restores things while also avoiding the interpolation of the array into the AST because it makes e.g. @code_warntype annoying to read. Also, make the array into a tuple because that tends to codegen slightly better.

Before:

julia> a = rand();

julia> @btime Base.Ryu.writeshortest($a)
  1.544 μs (40 allocations: 1.02 KiB)
"0.7946694032978381"

after

julia> @btime Base.Ryu.writeshortest($a)
  49.430 ns (2 allocations: 432 bytes)
"0.5526067771685965"