[WIP] Make FastDEQs fast again

[avik-pal]

Main Changes:

• Moves from Flux to Lux -- Mainly to do away with the weird destructure/restructure and the several issues that come with it
• Drop dependencies on unreleased branches (still requires some upstream merges but after this PR we can finally register the package)
• Termination and Time-Stepping Tolerances have separate arguments

TODOs

• Work on the Discrete Solvers
• Jacobian Stabilization is completely broken
• Update docs

Fixes #37

Requires LuxDL/Lux.jl#7 & avik-pal/FluxMPI.jl#15

[avik-pal]

Note that the models used Tsit5 before and VCABM3 now which contributes to a minor improvement.

Original Numbers

• CIFAR10: 6 processes across 2 V100s -- 450s per epoch

Some numbers after this partial rewrite:

• CIFAR10: 1 process 1 V100 -- 320s per epoch
• CIFAR10: 2 processes 2 V100s -- 250s per epoch
• CIFAR10: 4 processes 4 V100s -- 120s per epoch

I might have messed up memory usage though, 4 processes on 2 GPUs error out sometimes 😢

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (main@b32e1b4). Click here to learn what that means.
The diff coverage is n/a.

❗ Current head 57a7abb differs from pull request most recent head 7e42c90. Consider uploading reports for the commit 7e42c90 to get more accurate results

@@           Coverage Diff           @@
##             main      #45   +/-   ##
=======================================
  Coverage        ?   85.71%           
=======================================
  Files           ?       14           
  Lines           ?      448           
  Branches        ?        0           
=======================================
  Hits            ?      384           
  Misses          ?       64           
  Partials        ?        0           

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more