Specialize map for eltype conversions of a ReshapedArray

[jishnub]

Previously,

julia> map(Float64, reshape(1:4, 2, 2))
2×2 Matrix{Float64}:
 1.0  3.0
 2.0  4.0

After this PR:

julia> map(Float64, reshape(1:4, 2, 2))
2×2 reshape(::StepRangeLen{Float64, Base.TwicePrecision{Float64}, Base.TwicePrecision{Float64}}, 2, 2) with eltype Float64:
 1.0  3.0
 2.0  4.0

The map is translated to the parent, which may use specialized methods defined for the parent. This might reduce allocations.

Previously

julia> @btime map(Float64, $(reshape(1:400, 20, 20)));
  1.054 μs (1 allocation: 3.25 KiB)

After this PR:

julia> @btime map(Float64, $(reshape(1:400, 20, 20)));
  264.575 ns (0 allocations: 0 bytes)

[ViralBShah]

Merge?

[jishnub]

The doctest that fails:

julia> x = reshape(1:12, 2, 3, 2)
2×3×2 reshape(::UnitRange{Int64}, 2, 3, 2) with eltype Int64:
[:, :, 1] =
 1  3  5
 2  4  6

[:, :, 2] =
 7   9  11
 8  10  12

julia> mask = map(ispow2, x)
2×3×2 reshape(::Vector{Bool}, 2, 3, 2) with eltype Bool:
[:, :, 1] =
 1  0  0
 1  1  0

[:, :, 2] =
 0  0  0
 1  0  0

whereas this used to be an Array on master. While this PR specifically introduced this change, I wonder if this will have performance impacts downstream? Some basic testing indicates a minor regression:

julia> @btime $x[$mask];
  742.000 ns (2 allocations: 144 bytes) # master
  888.705 ns (2 allocations: 144 bytes) # PR