PBS memory scaling factor leads to too many cores

[dsschult]

On a cluster with these settings:

scheduler = PBS
mem_per_core = 3500
max_memory_per_job = 14000
max_cpus_per_job = 4 

We can get the situation where 14000 MB of memory is requested, which is allowed, and then because of the 1.05 scaling factor (in write_submit_file()) for memory we get 5 cpus allocated, which is a problem.

Fix this to maintain the maximum cpus and memory requested, even if we have to reject the job.