You can borrow missing (to count of 8) bytes from adjacent value and then throw them away with post-processing, applying a mask. Looks like it's faster to stay "numeric" at all stages if possible. The "unpack-pad-pack-unpack", (tybalt89's solution) is somewhere in the middle in terms of performance. Not that you said speed is the goal, but anyway:

use strict;
use warnings;
no warnings 'portable';
use Benchmark 'cmpthese';

my $bytes_per_value = 5;
my $count = 1_000;

my $value = 0xf_dead_beef_4;
my $bin_value = substr (pack ('Q', $value), 0, $bytes_per_value);
my $buffer = $bin_value x $count;

my $fmt = sprintf "(b%d)*", $bytes_per_value << 3;

my $pad_len = 8 - $bytes_per_value;
my $mask = ~ 0 >> $pad_len * 8;

cmpthese -1, {
    strings => sub {
        my @values = map { oct '0b'.reverse ($_)} 
            unpack ($fmt, $buffer);
        return \@values;
    },
    numbers => sub {
        my @values = unpack "(QX$pad_len)$count", 
            $buffer . "\0" x $pad_len;
        $_ &= $mask for @values;
        return \@values;
    },
};        

__END__

           Rate strings numbers
strings  2386/s      --    -77%
numbers 10274/s    331%      --
[download]