Add configuration section for global feature parameters

[arnavgarg1]

This PR enables users to specify type-specific default preprocessing, encoder, decoder and loss related parameters by introducing a new top-level config section called defaults.

In particular, this PR:

1. Moves existing feature preprocessing globals out of the preprocessing section such that the preprocessing section is only for things like splitting, etc.
2. Add a new defaults section that covers feature type defaults across all features.

Example:

preprocessing:
    split:
        type: random
        probabilities: [0.7, 0.1, 0.2]
    oversample_minority: 0.5
defaults:
    category:
        preprocessing:        
            missing_value_strategy: fill_with_const
            fill_value: <UNK>
        encoder:
            type: sparse
        decoder:
            norm_params: null
            dropout: 0
            use_bias: true
        loss:
            type: softmax_cross_entropy
            confidence_penalty: 0

[github-actions]

Unit Test Results

       5 files   -     1         5 suites   - 1   2h 10m 33s ⏱️ - 21m 28s
2 943 tests +    8  2 894 ✔️ +  8    49 💤 ±  0  0 ❌ ±0 
8 704 runs   - 101  8 542 ✔️  - 80  162 💤  - 21  0 ❌ ±0 

Results for commit 2979cad. ± Comparison against base commit 640821a.

♻️ This comment has been updated with latest results.

[justinxzhao]

Shouldn't we delete/update the merge_preprocessing function in preprocessing.py?

[arnavgarg1]

@justinxzhao I don't think that's necessary because https://github.com/ludwig-ai/ludwig/pull/2208/files#diff-671042bc7390c5a898a22f285e44a717dc6276bc0fd3e81142e8fc468e2e6e23R35 function does the work to pass in the data just like how it was being done before, which I also think is the cleanest way to do this.

This _merge_preprocessing is unnecessary and I've updated the code with all that's needed for this test.

[justinxzhao]

Just one comment, otherwise LGTM.

[justinxzhao]

@arnavgarg1 could you also take a stab at updating the documentation for 0.6? https://ludwig-ai.github.io/ludwig-docs/0.5/configuration/

[arnavgarg1]

@justinxzhao One thing I noticed from Connor's PR is that he's updated the tests to reflect the new nesting of encoders and decoders. Right now everything works because of backward compatibility, but should I create a follow up PR that modifies the tests to use the updated Ludwig config with defaults?

[arnavgarg1]

@justinxzhao One thing I noticed from Connor's PR is that he's updated the tests to reflect the new nesting of encoders and decoders. Right now everything works because of backward compatibility, but should I create a follow up PR that modifies the tests to use the updated Ludwig config with defaults?

Talked to @justinxzhao - will create a follow up PR with modified tests. Closing this issue for now.