mkimg: Indicate that input image pages are unlikely to be reused

[markj]

Summary

Currently, when mkimg(1) is used to create an image that's large
relative to system memory, it pulls its input files (e.g. filesystem
images) into the active queue in image_copyout_file(). This triggers
some aggressive behaviour from the pagedaemon, since it results in a
state where the inactive queue is depleted and almost all of system
memory is in the active queue. In particular, this forces the pagedaemon
to swap out lots of anonymous memory in an attempt to address the
shortfall.

Use madvise(2) in image_file_unmap() to hint to the VM that pages
backing input files should be reclaimed before other pages. This
basically means that they'll be reclaimed in FIFO order from the
inactive queue when the system is under memory pressure, avoiding
a global impact on the system.

Test Plan

Tested on a desktop-class system with 16GB of memory, creating
16GB raw bhyve images. With this change, I see an order of magnitude
less swap usage after creating an image. I also see a large runtime
improvement if most of the system's dirty anonymous memory has
not already been swapped out: 50s down to 20s. Once most inactive
pages have been swapped out, there's no significant difference in
runtime.

[marcel]

Very nice!