Relax type annotations to allow for non-array inputs

[adrhill]

The XAIBase interface currently requires inputs to be AbstractArray{<:Real}:

XAIBase.jl/src/analyze.jl

Lines 20 to 21 in 569675b

	function analyze(
	input::AbstractArray{<:Real},

This is too strict for explanations on e.g. named tuples (required by Transformers.jl, see Julia-XAI/ExplainableAI.jl#164) or tabular data. Ideally, input should be left untyped.

Example

julia> using ExplainableAI

julia> x = (foo=fill(1, 2, 2), bar=fill(2, 3, 3))
(foo = [1 1; 1 1], bar = [2 2 2; 2 2 2; 2 2 2])

julia> mymodel(x) = x.foo
mymodel (generic function with 1 method)

julia> mymodel(x)
2×2 Matrix{Int64}:
 1  1
 1  1

julia> analyzer = Gradient(mymodel)
Gradient{typeof(mymodel)}(mymodel)

julia> expl = analyzer(x)
ERROR: MethodError: no method matching (::Gradient{typeof(mymodel)})(::@NamedTuple{foo::Matrix{Int64}, bar::Matrix{Int64}})

Closest candidates are:
  (::Gradient)(::Any, ::AbstractOutputSelector)
   @ ExplainableAI ~/.julia/packages/ExplainableAI/rEdqm/src/gradient.jl:22
  (::AbstractXAIMethod)(::AbstractArray{<:Real}, ::Union{Integer, Tuple{Integer}}; kwargs...)
   @ XAIBase ~/.julia/packages/XAIBase/7QtIN/src/analyze.jl:33
  (::AbstractXAIMethod)(::AbstractArray{<:Real}; kwargs...)
   @ XAIBase ~/.julia/packages/XAIBase/7QtIN/src/analyze.jl:40

Stacktrace:
 [1] top-level scope
   @ REPL[11]:1