Correctness issues + annoying behavior

Hey,

The following behaviors are really annoying:

julia> t = TaylorScalar{Float64, 3}((Inf, 1.0, 0.0))     
TaylorScalar{Float64, 3}((Inf, 1.0, 0.0))

julia> t+1
TaylorScalar{Float64, 3}((Inf, 1.0, 0.0)) ##### OK

julia> t*1
TaylorScalar{Float64, 3}((Inf, NaN, NaN)) ##### Should be Inf, 1.0, 0.0

julia> t*2
TaylorScalar{Float64, 3}((Inf, NaN, NaN)) ##### Should be Inf, 2.0, 0.0

julia> 

The last one even looks like a correctness issue...

but also :

julia> t2 = TaylorScalar{Float64, 3}((0.0, 1.0, 0.0))    
TaylorScalar{Float64, 3}((0.0, 1.0, 0.0))

julia> t2^1
TaylorScalar{Float64, 3}((0.0, NaN, NaN)) ##### Should be 0.0, 1.0, 0.0

julia> t2^2
TaylorScalar{Float64, 3}((0.0, NaN, NaN)) ##### Should not be NaNs...

julia> 

Using TaylorSeries.jl dos not produce these NaNs, and this makes my test cases fail... Do you think this is something that would be fixable in the current state of TaylorDiff ?

Start by reproducing the TaylorScalar multiplication and power examples in the issue, then inspect the implementations of TaylorScalar arithmetic and exponentiation. Compare the results with TaylorSeries.jl and add coverage for the Inf and zero-base cases; done means the shown operations produce the stated coefficients without NaNs.