Set benchmark-specific thread limits for the Spark local executor

[lbulej]

The limits reflect the recommendations from #260, with the limits for gauss-mix, chi-square and dec-tree bumped up by 2 threads (providing a bit of slack that does not help but should not hurt), resulting in the following defaults for the spark_thread_limit parameter:

• 4 in chi-square and gauss-mix
• 6 in dec-tree
• 8 in movie-lens
• 12 in als, als-ml, log-regression and page-rank
• unlimited in naive-bayes

[vhotspur]

Related to #283: perhaps there should be some hard minimum too?

[lbulej]

Spark should be fine even with 1 CPU. On my machine (32G RAM, 4*2 CPUs), Spark benchmarks run (albeit slowly) under taskset -c 0, both on JDK8 and JDK11.

[lbulej]

If the thread limits are fine, I'd suggest merging this, because the issue with #283 appears to be related to thread stack size. If we end up imposing a minimal number of threads to work around the thread stack size, it would be better to do it in another PR anyway.