Reconsider ebpf map size for off-cpu profiling

[rockdaboot]

Summary

This has been raised at #316 (comment)

Isn't this sizing excessive if we bound it by number of possible PIDs? For example,
if we set off-cpu-threshold to 1000, we'll end up with a map that takes up roughly
4194304*16 = 64MB since we pre-allocate.

Generated off-cpu events are bounded by cores and scheduler call frequency, so for example
we could size this map as: next_pow2(cores*1000*pollInterval) using 1000Hz as an upper
bound estimate of scheduler call frequency which should consume far less memory.