Allow withgradient to pass aux info, and add @grad dropout

[mcabbott]

This upgrades withgradient to match the Zygote one:

julia> using Tracker

julia> withgradient(nt, 2) do x, p
         (; loss = sum(abs2, x.vec) ^ p, aux = x.vec)
       end
(val = (loss = 25.0, aux = [1.0, 2.0]), grad = ((vec = [20.0, 40.0], mat = [0.0;;], fun = nothing), nothing))

And it adds a rule for NNlib.dropout, which previously gave an error.

(These are united only by trying to make Flux work with Tracker again.)

[codecov]

Codecov Report

Attention: Patch coverage is 85.71429% with 3 lines in your changes are missing coverage. Please review.

Project coverage is 73.10%. Comparing base (bff6a0c) to head (c6ed3b1).

Files	Patch %	Lines
src/lib/array.jl	80.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #162      +/-   ##
==========================================
- Coverage   73.30%   73.10%   -0.21%     
==========================================
  Files           9        9              
  Lines         708      725      +17     
==========================================
+ Hits          519      530      +11     
- Misses        189      195       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[coveralls]

Pull Request Test Coverage Report for Build 8123435238

Details

• 18 of 21 (85.71%) changed or added relevant lines in 2 files are covered.
• 4 unchanged lines in 1 file lost coverage.
• Overall coverage decreased (-0.2%) to 73.103%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/lib/array.jl	12	15	80.0%

Files with Coverage Reduction	New Missed Lines	%
src/lib/array.jl	4	67.02%

Totals
Change from base Build 7535386303:	-0.2%
Covered Lines:	530
Relevant Lines:	725

---

💛 - Coveralls