[SPARK-10736] [ML] Use 1 for all ratings if $(ratingCol) = ""

[yanboliang]

For some implicit dataset, ratings may not exist in the training data. In this case, we can assume all observed pairs to be positive and treat their ratings as 1. This should happen when users set ratingCol to an empty string.

[SparkQA]

Test build #43078 has finished for PR 8937 at commit 5643680.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[mengxr]

It might be easier to define the rating col first:

val r = if ($(ratingCol) != "") col($(ratingCol)).cast(FloatType) else lit(1.0f)
val ratings = dataset.select(..., ..., r).map { row =>
  ...
}

[SparkQA]

Test build #43115 has finished for PR 8937 at commit 9162a22.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[mengxr]

LGTM. Merged into master. Thanks!