Fix job completion with big job queues

Accidentally stumbled upon this while testing unrelated code on a
machine with ~3K active jobs - the bash completion unittest was
hanging.

Upon investigation, it turns out that bash's ${var//pattern/repl/} is
probably quadratic in the size of input (or worse, even):

  $ touch job-{1..500}
  $ time ( a=$(echo job-*); echo ${a//job-/}| wc -c; )
  1892

  real    0m0.597s
  user    0m0.590s

  $ touch job-{1..1000}
  $ time ( a=$(echo job-*); echo ${a//job-/}| wc -c; )
  3893

  real    0m4.654s
  user    0m4.580s

We can easily fix this if we change to array-based substitution (once
per element):

  $ time ( a=($(echo job-*)); echo ${a[*]/job-/} |wc -c; )
  3893

  real    0m0.028s
  user    0m0.010s

  $ touch job-{1..10000}
  $ time ( a=($(echo job-*)); echo ${a[*]/job-/} |wc -c; )
  48894

  real    0m0.233s
  user    0m0.220s

This means that exactly when the master node is busy processing many
jobs, we could accidentally start consuming lots of CPU in the bash
completion, which is not good.

Note: the code might have problems with filenames containing spaces (I
didn't reset the IFS, etc.), but the original code had the same issue,
I think.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>