Specialize _memory_offset on number of arguments.

[chriselrod]

using BenchmarkTools
N = 5; data4big = randn(N,N,N,N,100); data4view = @view data4big[:,:,:,:,5];
@benchmark pointer($data4view)

Before:

julia> @benchmark pointer($data4view)
BenchmarkTools.Trial:
  memory estimate:  848 bytes
  allocs estimate:  21
  --------------
  minimum time:     1.480 μs (0.00% GC)
  median time:      1.562 μs (0.00% GC)
  mean time:        1.654 μs (1.72% GC)
  maximum time:     288.553 μs (98.46% GC)
  --------------
  samples:          10000
  evals/sample:     10

After:

julia> @benchmark pointer($data4view)
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     2.658 ns (0.00% GC)
  median time:      2.672 ns (0.00% GC)
  mean time:        2.686 ns (0.00% GC)
  maximum time:     12.688 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000

[KristofferC]

Seems like a pretty bad performance bug so I think it makes sense to backport this