Type dependent default maximum chunk size.

From the following benchmark

function rosenbrock(x)
   a = one(eltype(x))
   b = 100 * a
   result = zero(eltype(x))
   for i in 1:length(x)-1
       result += (a - x[i])^2 + b*(x[i+1] - x[i]^2)^2
   end
   return result
end

using BenchmarkTools
import ForwardDiff
for T in (Float32, Float64)
    println("T = $T")
    x = rand(T, 64)
    y = similar(x)
    for chunk in (2, 4, 8, 16, 32)
        print("Chunk: $chunk -")
        gradient_cache = ForwardDiff.GradientConfig(rosenbrock, x, ForwardDiff.Chunk{chunk}())
        @btime ForwardDiff.gradient!($y, rosenbrock, $x, $gradient_cache)
    end
end

with the result

T = Float32
Chunk: 2 -  10.411 μs (0 allocations: 0 bytes)
Chunk: 4 -  6.778 μs (0 allocations: 0 bytes)
Chunk: 8 -  7.124 μs (0 allocations: 0 bytes)
Chunk: 16 -  4.576 μs (0 allocations: 0 bytes)
Chunk: 32 -  3.200 μs (0 allocations: 0 bytes)
T = Float64
Chunk: 2 -  9.029 μs (0 allocations: 0 bytes)
Chunk: 4 -  7.851 μs (0 allocations: 0 bytes)
Chunk: 8 -  5.353 μs (0 allocations: 0 bytes)
Chunk: 16 -  5.005 μs (0 allocations: 0 bytes)
Chunk: 32 -  4.351 μs (0 allocations: 0 bytes)

it suggests that maybe the maxumum chunk size should be a function of the type used.

Another thing to consider is that perhaps we should try use chunk size which well into vector registries (assuming they don't cause another function evaluation)

t = Float32
println("T = $T")
x = rand(T, 21)
y = similar(x)
for chunk in (8, 10)
    print("Chunk: $chunk -")
    gradient_cache = ForwardDiff.GradientConfig(rosenbrock, x, ForwardDiff.Chunk{chunk}())
    @btime ForwardDiff.gradient!($y, rosenbrock, $x, $gradient_cache)
end

with the result

Chunk: 8 -  13.837 μs (0 allocations: 0 bytes)
Chunk: 10 -  15.309 μs (0 allocations: 0 bytes)

so 8 here is a better choice than 10 (fitting nicely into an AVX lane).

Start by reproducing the provided BenchmarkTools examples for Float32 and Float64, focusing on ForwardDiff.GradientConfig, ForwardDiff.Chunk, and gradient!. Compare chunk sizes and vector-register effects, then determine a type-dependent default and verify that the benchmark shows the intended performance improvement.