Polynomial performance improvements

[Joel-Dahne]

This adds a number of minor improvements to polynomials, mostly related to construction.

The main goal is to improve the performance when constructing ArbSeries with of the form ArbSeries([x, 1]), mostly by reducing the number of allocations. It makes the following changes related to this

1. Allow passing tuples when constructing polynomials, saving one allocation in constructing the vector. So now ArbSeries((x, 1)) works. This change is breaking in the sense that previously passing a two-tuple (a, b) would set the first coefficient to Arb((a, b)), i.e. the interval from a to b. This could now be done with ArbSeries(((x, 1),)).
2. Make setindex! work well with Int using arb_poly_set_coeff_si. Previously this would convert the input to Arb before setting the coefficient.
3. Make a specialized constructor for two-tuples optimized for tuples with heterogeneous types. If you naively iterate over the elements in the tuple with a for loop the type inference doesn't work and you suffer in performance. By explicitly unfolding the loop type inference works well.
4. I have also removed the fit_length! in one place in the construction for series, this is already done by the underlying constructor so doesn't have to be done again.

Before the above changes the fastest way to construct an ArbSeries with coefficients [x, 1] was with ArbSeries([x, one(x)]), now it is ArbSeries((x, 1)), comparing the performance we see that the later is much faster and (more importantly) allocates a lot less.

julia> x = Arb(1)
1.0000000000000000000000000000000000000000000000000000000000000000000000000000

julia> @btime ArbSeries([$x, one($x)])
  192.790 ns (4 allocations: 320 bytes)
1.0000000000000000000000000000000000000000000000000000000000000000000000000000 + 1.0000000000000000000000000000000000000000000000000000000000000000000000000000⋅x + 𝒪(x^2)

julia> @btime ArbSeries(($x, 1))
  112.588 ns (2 allocations: 144 bytes)
1.0000000000000000000000000000000000000000000000000000000000000000000000000000 + 1.0000000000000000000000000000000000000000000000000000000000000000000000000000⋅x + 𝒪(x^2)

In addition to the above I added @inbounds in one place where it was missing and better documentation for Arblib.ref for polynomials.

Since supporting constructing polynomials from tuples is a breaking change I have pushed the version to 0.6.0.

[kalmarek]

This looks good.

Probably we should also support a vararg constructor as well, but this may wait for another time.

[Joel-Dahne]

I took the liberty of adding the bump ofSpecialFunctions.jl to this pull request from #143, it should not affect anything.