fix(cpworker): count every output packet once, with its sliced length

[timmy21]

Problem

The output counters (fwd_*, direction_drop_*, ratelimit_drop_*, error_drop_*) had no definition, and each output counted them differently (#308):

• gre/vxlan added the GRE/VXLAN header to fwd, error_drop and ratelimit_drop bytes, but not to direction_drop bytes.
• zmq counted fwd/error_drop bytes as whole batches, including the batch header, record headers and MPLS labels. Its direction_drop/ratelimit_drop bytes included the MPLS label.
• A gre/vxlan partial send counted the packet in fwd_packets and split its bytes across fwd_bytes and error_drop_bytes.
• A VXLAN-split packet counted once per fragment. A failed build_fragment returned success without counting the packet.
• rate_limit_mbps charged a different length per output, so fwd_bytes could not be compared with the limit.

Change

Each packet handed to an output is counted once, as 1 packet and L bytes, in exactly one bucket. L is the packet's caplen after slice, without any encapsulation.

Output	Change
gre	All buckets and the rate limiter use L (no GRE header). A partial send counts as error_drop only
vxlan	do_send_packet only sends; vxlan_send_packet counts once after the whole packet, all fragments included, is sent or fails. A partial send, a failed fragment or a failed fragment build counts as error_drop. All buckets and the rate limiter use L
zmq	The batch tracks data_bytes, the frame bytes of its captured packets (replacing heartbeat_bytes); a flush adds it to fwd_bytes or error_drop_bytes. Direction/ratelimit drops and the rate limiter no longer add the MPLS label
null, file, rotating_file	Already followed the definition; unchanged

Naming (second commit): length/caplen meant a different length in each output. output_frame_len() in output.h computes the counted length (caplen after slice, capped per output: GRE_MAX_FRAME_LEN, VXLAN_MAX_FRAME_LEN, ZMQ_MAX_FRAME_LEN), and gre, vxlan, null and zmq call it frame_len. header->caplen always means the captured length; vxlan uses payload_len (frame + capture_time), zmq record_len (frame + MPLS label). send_whole_or_split takes only frame_len, so a split packet is parsed with the same length it is sent with (they differed when caplen > 65535).

Docs: the definition is in docs/REFACTOR-CPCTL.md §6.2. ZMQ-WIRE-FORMAT §8 and the rate_limit_mbps row in USAGE-CPWORKER refer to it (English and zh-Hans).

Behaviour changes

• gre, vxlan and zmq report lower fwd_bytes/drop bytes than before: encapsulation headers are no longer counted. These values are uploaded to CPM as fwdBytes.
• vxlan with split reports one fwd_packets per original packet instead of one per fragment.
• rate_limit_mbps for gre, vxlan and zmq is slightly looser, since the 4–8 byte headers are no longer charged. While the limit is active, fwd_bytes grows by at most rate_limit_mbps.

Tests

• New tests/unit/output_vxlan.c: sliced length without VXLAN header or capture_time, split counted once, send error (EMSGSIZE) counted as one error_drop, rate limit charging L. A GRE rate-limit/direction test runs when a raw socket is available and is ignored otherwise.
• tests/unit/output_zmq.c: two assertions now expect frame bytes; new tests for sliced length plus conservation, and for the rate limit charging L.
• The new tests fail against 0.9.x (vxlan 5/5, zmq 4).
• pktm VM, on both commits: make test 12/12 passed (GRE test also run under sudo); run_test.sh all passed.

Fixes #308