Impact: Reduce kernel memory usage, use new cpumask API.

Avoid a static cpumask_t for prof_cpu_mask, and an on-stack cpumask_t
in prof_cpu_mask_write_proc.  Both become cpumask_var_t.

prof_cpu_mask is only allocated when profiling is on, but the NULL
checks are optimized out by gcc for the !CPUMASK_OFFSTACK case.

Also removed some strange and unnecessary casts.
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>