### What changes were proposed in this pull request?
availableSlots are computed before the for loop looping over all TaskSets in resourceOffers. But the number of slots changes in every iteration, as in every iteration these slots are taken. The number of available slots checked by a barrier task set has therefore to be recomputed in every iteration from availableCpus.
### Why are the changes needed?
Bugfix.
This could make resourceOffer attempt to start a barrier task set, even though it has not enough slots available. That would then be caught by the `require` in line 519, which will throw an exception, which will get caught and ignored by Dispatcher's MessageLoop, so nothing terrible would happen, but the exception would prevent resourceOffers from considering further TaskSets.
Note that launching the barrier TaskSet can still fail if other requirements are not satisfied, and still can be rolled-back by throwing exception in this `require`. Handling it more gracefully remains a TODO in SPARK-24818, but this fix at least should resolve the situation when it's unable to launch because of insufficient slots.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Added UT
Closes#23375Closes#25946 from juliuszsompolski/SPARK-29263.
Authored-by: Juliusz Sompolski <julek@databricks.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
420abb457d