Better BigFloat support

[ErikQQY]

Continue #438 and SciML/BoundaryValueDiffEq.jl#202

When handling with BigFloat u0, instead of specializing when creating a new array, we directly fill zeros everywhere.

[avik-pal]

This will double the allocations once for similar and again a new zero array.

Also performance-wise this is not a good idea: (exaggerated example but still)

julia> x = rand(100000);

julia> @btime similar($x);
  108.556 ns (3 allocations: 781.31 KiB)

julia> @btime zero($x);
  96.660 μs (3 allocations: 781.31 KiB)

[avik-pal]

https://github.com/SciML/SciMLStyle#always-default-to-constructs-which-initialize-data says that it is okay to use undef if we ensure that the data is initialized before being returned, so I don't think it applies in this context.

[ChrisRackauckas]

Where are you zeroing it?

[avik-pal]

The u in solution object will be initialized right? If the user doesn't assign a value to an element of u inside the in place function that is also an undefined behavior.