Add simple RNN functionality

[drhuffman12]

Add simple RNN functionality (See Issue: #15 )

• Phase 1: initial prototype/skeleton (w/ basic structure calc's) (See PR: Drhuffman12/cmn basic rnn (part 1) #46 )
• Phase 2: expand prototype/skeleton (to include the associated 'mini_net' objects) (See PR: Drhuffman12/cmn basic rnn (part 2) #47 )
• Phase 3: expand prototype/skeleton (to include the methods to 'guess' and 'train') (See PR: Drhuffman12/cmn basic rnn (part 2) #47 )
• Phase 4: FIX to/from json (use type key?) (in a later PR, at least partly in PR: Drhuffman12/cmn basic rnn (part 2) #47)
• Add code to: (See PR: drhuffman12/cmn_basic_rnn_part_5 #51 )
  • Collect (split?) a 'sequence of input and output data' into small chunks for use by RnnSimple#train(..)
  • Add a variation of RnnSimple#train(..) that loops thru the 'sequence of input and output data'
• add initial BreederUtils and Breeder (See PR: Drhuffman12/cmn basic rnn part 6 #54)
• Consolidate error_distance related code into ErrorStats (See PR: drhuffman12/add_team_utils (part 1) #56 and https://github.com/crystal-lang/crystal/tree/master/.github/workflows)
• Convert misc classes to use ErrorStats (See PR: drhuffman12/add_team_utils (part 1) #56)
• Reorg/refactor props and inits (See PR: Drhuffman12/add team utils part 2 #57)
• Add a net mixer
  • (?) ~~BreedParent
  • Breeder class(es) (renamed to *Manager classes)
    • Abstract Breeder(T) (See PR: drhuffman12/add_team_utils (part 1) #56)
    • MiniNetBreeder < Breeder(MiniNet) (See PR: Drhuffman12/add team utils part 2 #57)
    • ChainBreeder < Breeder(Chain)
    • RnnSimpleBreeder < Breeder(RnnSimple) (See PR: Drhuffman12/add team utils part 3 #58)
    • (?) BackproagationBreeder < Breeder(Backproagation)
    • Add misc methods for picking delta (See PR: Drhuffman12/add team utils part 2 #57)
      • random (default)
      • based on some mix of both parent's error scores (ideal)
    • switch to using 'CounterSafe'
  • (?) TeamUtils (See PR: Drhuffman12/add team utils part 4 #59)
    • Should mix misc combo's of nets
    • Should use delta based on some mix of both parent's error scores
    • re MiniNetManager
    • re RnnSimpleManager
  • Mod app and test code in spec_bench/ai4cr/neural_network/rnn/rnn_simple_manager_spec.cr re training a team of RNN nets on a text file.
    • See: examples/rnn_simple_manager_example.cr
    • Mod text util so that when converting float bits to chars, it uses:
      (a) 0.0 when <= 0.0
      (b) 1.0 when > =1.0
      (c) rounds bits [to 0.0 or 1.0]
    • Add #(un)certainty methods for iod data
    • Add code/tests/benches for utf text files (for RNN usage)
    • test: errors should decrease (kinda depends on net and training data structure/size)
    • update this app and shards to be Crystal v1.0 compatible
• Implement Bi-directional RNN (i.e.: RnnSimple pulls from inputs and previous time column.)
• Switch to (or add associated classes to use) BigFloat instead of Float64
  • Reasons:
    (a) Outputs are getting too big when using RELU for larger networks.
    (b) I tried handling nan and infinity values via forcing to 0,1,etc or by breeding and purging, but that starts to fail for larger networks.
    (c) I tried auto-scaling down the initial weights (based on network params) to avoid (a), but when scaled down too much (i.e.: larger networks), I get arithmetic errors.
    (d) The things I have tried (or upgrade to Crystal 1.0) seems to now tend to lead to more memory usage (sometimes maxing out my memory).
• Refactor Chain?); e.g.:
• Fix spec_bench (re Chain?); e.g.:

  --------
    plot: ''
    error_stats.history: '[]'
  

• CI should test builds for:
  • Linux
    • Ubuntu (x86) (from the beginning!)
    • Alpine (aarch64)
  • Mac
  • Windows
  • See: https://github.com/crystal-lang/shards/blob/master/.github/workflows/ci.yml
• Update https://drhuffman12.github.io/ai4cr/
• (?) Phase 5: Add more examples and benchmarking (e.g.: simple wave form learning)
• WHY is the last ti's value in outputs_guessed reset to '0.0' after training (but NOT after eval'ing)??? (and NOT reset to '0.0' after next round of training???)
• (?) Convert RnnSimple's validation @errors to Hash(<Enum>, String). (I switched it from Hash(Symbol, String) to Hash(String, String), due to issues w/ #from_json. An Enum would probably be more efficient than a String.)
• Convert all specs to Spectator format.
• Likewise, split up Chain into chain_concerns and fix where applicable.
  • Check for cleanup in app and tests
• (re)Update docs (written to https://drhuffman12.github.io/ai4cr/)
• Code cleanup!
• Update version!!!