This change modifies vm_page_grab_pages to use vm_page_alloc_pages.
Instead of allocating and inserting one page at a time,
vm_page_grab_pages will now inspect the object's memq or the radix
tree to determine how many pages can be allocated from the
current offset and allocate them using vm_page_alloc_pages.

Tested by: gallatin

All changes have been stress-tested and performance-tested using sendfile.
@gallatin has tested the entire patch series on a machine with the following specs:
Xeon E5-2697A v4 @ 2.60GHz, 1x Mellanox ConnectX4 100GbE, 4x nvme, 128GB DDR4-2400
The results showed a 3.5% reduction in CPU usage when serving Netflix video traffic.

I've also evaluated vm_page_alloc_pages using a simple microbenchmark that allocates
N pages at a time and compared it to the common 'vm_page_alloc_page + forloop' approach.
The plotted results are available at https://people.freebsd.org/~bnovkov/plots/batch_alloc_plot.png