Support efficient hashing of sparse files

[frankdavid]

b3sum is quite slow when operating on sparse files.

To reproduce, create a sparse file with some data and a hole and invoke b3sum:

$ dd bs=1G count=1 if=/dev/random of=/tmp/sparse.dat
$ truncate -s=30G /tmp/sparse.dat

$ ls -lsh /tmp/sparse.dat
# Observe that the file size is 30G but only 1.1G on disk

$ time b3sum /tmp/sparse.dat

real	0m22.195s
user	0m12.049s
sys	3m46.161s

Solution:

1. Linux provides a way using lseek with SEEK_DATA+SEEK_HOLE to identify where the data resides in a sparse file. This can be used to optimize the implementation.
2. Before processing a file, we collect the list of data segments using lseek.
3. While we process the input, we use the list of data segments to know whether we are in a data segment or a hole segment.
4. When we are in a hole segment, instead of reading from the mmap-ed input, we read from a static zero array of the same size. The benefit of this is that reading from the static zero array does not cause a page fault and a context switch. Furthermore, the static zero array will likely be in L1 cache.

The effect of this optimization if quite significant:

real	0m0.667s
user	0m9.013s
sys	0m0.228s

The POC can be found here. If you're happy with this direction, I'll still refactor the code slightly and add tests.