fix some cases of gradient/jacobian with StaticArrays

[ExpandingMan]

This fixes at least some cases of functions taking and return static arrays being returned as Array (and the associated allocations).

An example of a case this fixes is

f1(x) = @SVector Float64[x[1], 2*x[2]]

x = @SVector ones(2)

gradient(Forward, f1, x)

There were two things wrong with the tupstack function in current main. The first is that cat always returns an Array. This may be a missing method in StaticArrays, but I suspect that the reason it's missing is that it's hard to statically infer the resulting shape (and it would rely on static inference of the value of dims). Suffice it to say, I think it's safe to assume that cat can never be used with staticarrays.

The second issue is that when reshape is used with static arrays, one must use Size arguments to return a static array, to indicate that the size can be statically inferred.

This works in the cases I have tested but looks fragile. I still need to confirm more thoroughly that this does not break any of the higher rank cases.

[wsmoses]

Add a test?