Fix out of bound access of the IPv6 address prefix if /128 is used.

Similarly to the helpers defined in tests/harness/filters.c, add a new
function to create a test set: bf_test_set_get().

`/usr/bin/env python` should not be used as genpkts.py hasn't been
tested with Python2, look for Python 3 instead.

Use find_package(Python3) to locate the Python 3 interpreter on the
system, intead of relying on genpkts.py's shebang. This way CMake will
warn early about a missing dependency.

This is a very specific set type expected to be used internally at Meta.
Eventually, generic sets will allow for any set key to be dynamically
generated.

Add BF_MATCHER_SRCIP6PORT to match against a set of (source IPv6, source
port) tuples set.

The BPF bytecode will ensure the packet contains an IPv6 and a TCP or
UDP header to filter on.

Tests the new BF_MATCHER_SRCIP6PORT matcher using 200k entries in a set.

Serialize all the element of a set in a single marsh, instead of
creating a marsh for every single element. The bigger the set the
greater the performance gain.

Tested on a 200k elements set with a performance gain of ~10000x.
Because the type of the set is stored in the serialized data, there is
no need for each element to have its own container.

bf_cgen_dump() has a negative performance impact even if nothing is
printed (as DUMP() uses bf_dbg() which is also gated to the verbose
mode).

Instead, only call bf_cgen_dump() if the verbose mode is enabled.

Use offsetof() instead of bf_btf_get_field_off() for ingress_ifindex
field of the xdp_md structure, this change saves us from searching for a
specific symbol name in the kernel BTF.

Use BPF_MAP_UPDATE_BATCH to fill the BPF maps used to define the
filtering sets, instead of inserting every element one after the other.

The child process blocks once the pipe buffers (stdout and stderr used in
the subprocess) are full, leading to confusing hangs. Set the buffers
size to 1MiB preemptively. This should be enough are the process
structures are used for testing.

If this issue resurfaces, more control should be provided to the user to
increase the buffers size or set the buffers as non blocking (or flush
the buffers).

The test harness could be used outside of bpfilter to run custom
benchmarks on bpfilter. Make the harness binary a static library.

Add support for sets containing source IPv6 addresses. This implement is
subpar but is required for benchmarking. Eventually, bpfilter will
support generic sets (user-defined sets with custom keys) and static set
types will be removed.