Drop field dimension, support linear indexing and empty fields

[charleskawczynski]

This is an alternative to #1929. I opened this just to preserve the previous branch/PR for convenience. This branch is probably more matured.

This CPU tests are nearly fully passing.

Remaining issues / items to revisit

• field_name now allocates
• allowed_max_eps_error changes (perhaps just due to copied data?)
• Remove field_array(array::AbstractArray, ::Type{T}) where {T <: AbstractData} if possible
• Remove DataLayout singletons (for constructing FieldArrays) if possible
• ) broken = Sys.iswindows() is this actually somehow fixed?
• column_field2array_view cannot work but column_field2array is fine. (column_field2array_view is only used in the test dir)
• Test cuda
• Avoid collect in writers: array = collect(parent(field)) # TODO: avoid collect here
• Revert to single_field_solver_cache(::UniformScaling, b) = similar(b, Tuple{}) if possible
• Leave column_field2array_view and use copies in test suite
• Equality of empty fields?
• RecursiveApply on fieldvectors (test/Fields/unit_field.jl)
• ForwardDiff support (test/Fields/unit_field.jl)
• Review remapping changes (src/Operators/remapping.jl)

[charleskawczynski]

Since they're working, here's the results for the CPU benchmarks:

                  Main                                  this PR
copyto! (DataF):  2 nanoseconds                     |   2 nanoseconds
copyto! (IJFH):   45 microseconds, 792 nanoseconds  |   14 microseconds, 333 nanoseconds
copyto! (IFH):    23 microseconds, 250 nanoseconds  |   3 microseconds, 510 nanoseconds
copyto! (VF):     8 nanoseconds                     |   7 nanoseconds
copyto! (VIJFH):  2 milliseconds, 219 microseconds  |   1 millisecond, 440 microseconds
copyto! (VIFH):   402 microseconds, 750 nanoseconds |   286 microseconds, 417 nanoseconds

for the following parameters:

FT = Float64
S = FT
Nf = 1
Nv = 63
Nij = 4
Nh = 30 * 30 * 6
Nk = 6

Some cuda benchmarks are already summarized in #1910 and here