Safe functional indexing

[Tokazama]

This allows common comparison operators be parsed into indices (<, <=, >, >=). So instead of doing x[3:end] you can do x[>=(3)] (if x uses ArrayInterface.getindex. The main utility right now is that we can ensure that the indices are always inbounds using this syntax. I'm still working out the best way to communicate this info to checkbounds so that bounds checking is passed over (similar to how : becomes Slice which is known to be inbounds); but that will probably be more involved so I'm going to make that a separate effort.

[codecov]

Codecov Report

Merging #316 (5ad45fd) into master (b922313) will increase coverage by 0.12%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #316      +/-   ##
==========================================
+ Coverage   91.74%   91.86%   +0.12%     
==========================================
  Files           9        9              
  Lines        1429     1439      +10     
==========================================
+ Hits         1311     1322      +11     
+ Misses        118      117       -1     

Impacted Files	Coverage Δ
src/ArrayInterface.jl	86.17% <100.00%> (+0.22%)	⬆️
src/indexing.jl	89.75% <100.00%> (+0.77%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b922313...5ad45fd. Read the comment docs.

[Tokazama]

@wangl-cc, you're welcome to review this since it may be of interest to you.

[wangl-cc]

If checkbounds is needed, I think min or max can be removed, like the syntax

julia> x = 1:2
1:2

julia> Meta.@lower x[begin:3]
:($(Expr(:thunk, CodeInfo(
    @ none within `top-level scope`
1 ─ %1 = Base.firstindex(x)
│   %2 = %1:3
│   %3 = Base.getindex(x, %2)
└──      return %3
))))

Besides, I think we can create some "functional range" like getindex(x, 3:last), which is much similar to the syntax x[3:end].

[Tokazama]

The reason I used min and max was because those functions are returning all indices below or above the provided value. This is the same as if you filtered the axis with those functions. Maybe this needs to be clarified more in the docs.

We can't create a functional range like that because it would be type piracy.

[wangl-cc]

The reason I used min and max was because those functions are returning all indices below or above the provided value. This is the same as if you filtered the axis with those functions. Maybe this needs to be clarified more in the docs.

With this interpretation, I think keeping them always inbounds is OK.

We can't create a functional range like that because it would be type piracy.

Maybe we can create new types as alternatives to begin and end to avoid type piracy.

[Tokazama]

There is a package for the last suggestion, so it's probably more intuitive to do something similar or provide support through that package. I'll remove the firstindex and lastindex ones for now.