[SPARK-8880] Fix confusing Stage.attemptId member variable

[kayousterhout]

@squito this seemed like something we should just fix, independent of your change. Let me know what you think of this fix.

[SparkQA]

Test build #36741 has finished for PR 7275 at commit e150278.

• This patch fails Scala style tests.
• This patch merges cleanly.
• This patch adds no public classes.

[squito]

Looks good (with the minor style fix) -- much clearer than it was before!

I was a little concerned about what would happen when you submit a job which re-uses a previously computed stage -- would the stage attempt ids inside the job start be consistent before and after this change? But I ran a small example and I realized it doesn't matter. If a shuffle output is reused, we dont' actually reuse the stage, we make a new one which gets skipped. (I thought there was some circumstance where two jobs could refer to the exact same stage, but I can't seem to recreate it now ... is it possible?)

Its kinda weird that job start even has the attempts for the stage, since they will always be 0, but I guess that is just an artifact of reusing the StageInfo object.

[squito]

needs a return type

[SparkQA]

Test build #37050 has finished for PR 7275 at commit 3e9ce7c.

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

[squito]

lgtm