Reduce the allocations of gausslegendre for moderate n and add a flag for Aqua

[HMegh]

Hi,

Reducing allocations:

The Newton's iteration for gausslegendre uses only part of the information returned by the asymptotic formula asy. I moved the relevant code from asy to a new function leg_initial_guess to reduce allocations and improve performance. I have also reused the temporary vectors PP1 and PP2. This reduces allocations in rec(n) from 22 to 9.

Example 1
Current Master:

julia> @btime gausslegendre(10)
  541.186 ns (22 allocations: 1.36 KiB)

With this PR:

julia> @btime gausslegendre(10)
  396.139 ns (9 allocations: 640 bytes)

This performance improvement is less pronounced for higher values of n since the Newton iteration is O(n^2).

Example 2
Current Master:

julia> @btime gausslegendre(50)
  10.787 μs (22 allocations: 5.52 KiB)

With this PR:

julia> @btime gausslegendre(50)
  10.379 μs (9 allocations: 2.62 KiB)

CI and Aqua

Currently the tests are not passing since actions/cache@v1 is deprecated, and Aqua.jl is indicating that the packages in extras do not have a compat entry. I have upgraded the actions/cache version, and added a flag to Aqua to ignore packages in extra. All tests pass now.

Best,
Haroun