add better implementations for generic findprev findnext for AbstractString

[KristofferC]

Ref #35727 (comment)

using DelimitedFiles
A = randn(Complex{Float64},300,50)
writedlm("A.csv",A,',')
@time readdlm("A.csv",',',Complex{Float64},'\n')

Before:

  9.316321 seconds (293.64 k allocations: 3.262 GiB, 3.35% gc time)

After

 0.013526 seconds (30.09 k allocations: 1.882 MiB)

Fixes #35721.

[timholy]

An almost 1000x improvement, not too shabby.

[stevengj]

Out of curiosity, how does the speed of findnext compare to the old version?

Update: On my machine with Julia 1.5 master, this findnext is about 20% faster at finding the last character of a 1000-character string or a 37% improvement for an 11-character string. So, the old iterator-based implementation is not bad (and has no allocation), but the new implementation is slightly faster (and arguably simpler).

(Of course, the biggest improvement is eliminating the allocation in findprev.)

[stevengj]

Note: Fixes #35721.

[KristofferC]

FreeBSD failure looks unrelated.