[SPARK-18434][ML][FOLLOWUP] Add checking for setSolver in GLM

[zhengruifeng]

What changes were proposed in this pull request?

add checking in GLM.setSolver to forbidden unsupported solvers:
val glm = new GeneralizedLinearRegression().setSolver("123")

How was this patch tested?

existing tests

cc @yanboliang

[zhengruifeng]

HasSolver now don't support value validation, so it's not easy to use:
GLR and LiR inherit HasSolver, but need to do param validation explicitly like this PR
MLP dont inherit HasSolver, and create param solver with supportedSolvers
It may be reasonable to modify HasSolver after 2.1

[SparkQA]

Test build #68918 has finished for PR 15955 at commit 943edba.

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

[yanboliang]

@zhengruifeng It's better we can define a variable supportedSolvers in object GeneralizedLinearRegression and check it contains users' input value. I think it also applies to LinearRegression, could you also change that? Thanks.

[zhengruifeng]

@yanboliang I updated this pr according to your comments. And now it's solver looks like MPLC

[SparkQA]

Test build #68977 has finished for PR 15955 at commit 1fe2f92.

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

[yanboliang]

One minor comment, otherwise, LGTM. Thanks.

[yanboliang]

Could we make it more private (i.e. private[GeneralizedLinearRegression])?

[yanboliang]

Ditto.

[zhengruifeng]

@yanboliang I add the private[regression] limitation refering to MultilayerPerceptronClassifier implementation: https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/ml/classification/MultilayerPerceptronClassifier.scala#L272

[zhengruifeng]

After offline discussion with @yanboliang , I will update this PR to make solvers more private and include MLPC in this PR.

[yanboliang]

After offline discussion with @zhengruifeng , we think the validation in the Param itself should be better, so other ways (which not calling setter method) of setting params can also benefit from the validation. We open SPARK-18518 to a more perfect solution.