[R-package] Improvements, readability, and bug fixes

[Laurae2]

This PR is large-sized changes for the R package. There are issues to solve which will not be fixed in this PR, as they are already on the master branch.

R improvements and bug fixes:

• All changes are (supposed to be) without code regression (I added only on top, or edited what we have by adding more elements)
• Allow to use regression_l1, regression_l2, huber, fair, poisson, which were unusable previously due to hard-coded rules.
• Auto-define default loss functions for regression_l1 (MAE), regression_l2 (MSE), huber (MAE), fair (MAE), poisson (Poisson loss):

• Changed the way of pre-allocating zero-valued vectors (slightly faster).
• Users who have xgboost library used after lightgbm can now again use lightgbm examples (getinfo, setinfo, slice global environment name clash).
• Add lgb.unloader which wipes LightGBM environment so we can avoid restarting R when an object gets stuck in memory for no apparent reason (like when training over training different things on same variables which have changed).
• lgb.unloader can fully wipe LightGBM objects in the specified environment (lgb.Booster, lgb.Dataset), and will not cause an error when lightgbm was already detached from the R environment.
• Added new example for lgb.unloader.
• Remove "free booster handle" message on Predictor types
• Fix dim.lgb.Dataset, dimnames.lgb.Dataset functions' example dontrun tag

R readability/stylistic changes:

• Commented the whole R code from scratch (everything is now commented to make editing easier for newcomers)
• Changed all single quotes to double quotes for defining characters (consistency fix)
• Fix all the spacing issues (code way easier to edit as we don't have to mash back+space button to get the code at the right place...)
• Improved readability of code by splitting functions into "paragraph" chunks (group of actions) to ease editing and spotting critical chunks
• Made demonstration codes easier to read

R issues to solve in future PRs, not now:

• Reverse the following or use Travis as alternative: do not run examples on devtools::check on any function currently (might reverse this later, in the future - we would need Travis setup).
• [Windows only (?)] Do not lock DLL in R when building library, because it does not allow to use devtools::check() to inspect the whole code which is essential for checking for code correctness (without having to run everything by hand). Potentially a regression from PR Support build self-contained R package.  #340 to answer issue Relative paths in R-package prevent source package build #339 (not sure though, it is a strange lock bug). This will not pass in CRAN in its current state as the lock does not allow to run CRAN tests (need to unload library properly) - we still have no CRAN release, therefore this is not a major issue currently.
• Get rid of library(lightgbm) in examples (not allowed without \{dontrun})
• Stop relying for some allocations (<<-) on the global environment (not easy to fix at all).
• Provide the user demo code to convert list of evaluation metrics to a matrix or data.table (data.table::rbindlist).

Tests performed:

• Run all R function examples
• Run extra examples, see below for Poisson loss example

Example for testing a new loss (ex: Poisson loss):

library(lightgbm)
data(agaricus.train, package = "lightgbm")
train <- agaricus.train
dtrain <- lgb.Dataset(train$data, label = train$label)
data(agaricus.test, package = "lightgbm")
test <- agaricus.test
dtest <- lgb.Dataset.create.valid(dtrain, test$data, label = test$label)
params <- list(objective = "poisson") # DO NOT SPECIFY METRIC, ONLY OBJECTIVE like `regression_l1`, `regression_l2`, `huber`, `fair`, `poisson`
valids <- list(test = dtest)
model <- lgb.train(params,
                   dtrain,
                   100,
                   valids,
                   min_data = 1,
                   learning_rate = 1,
                   early_stopping_rounds = 10)

Beginning of training log (does not converge because it is not the right objective obviously, and there are too many 0s, which makes it nearly impossible - this is normal because it is not a dataset for Poisson regression):

[LightGBM] [Info] Total Bins 232
[LightGBM] [Info] Number of data: 6513, number of used features: 116
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Trained a tree with leaves=14 and max_depth=6
[1]:	test's poisson:0.523205 
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Trained a tree with leaves=24 and max_depth=8
[2]:	test's poisson:0.482608 
[LightGBM] [Info] No further splits with positive gain, best gain: -inf
[LightGBM] [Info] Trained a tree with leaves=24 and max_depth=7

[msftclas]

@Laurae2,
Thanks for having already signed the Contribution License Agreement. Your agreement was validated by Microsoft. We will now review your pull request.
Thanks,
Microsoft Pull Request Bot