Fix latency increase when the application idles between RPCs

[Stuart0l]

Fixes #116 (and #104). Root-cause analysis is in #116.

Changes

1. Bound the timing wheel's catch-up (src/cc/timing_wheel.h)

Added a num_wheel_entries_ counter to TimingWheel, maintained in insert_into_wslot() and reap_wslot(). reap() uses it to detect an empty wheel: when the wheel is empty and the lag exceeds one full revolution, it re-bases all slot timestamps in a single pass and returns, capping the cost at kWheelNumWslots iterations. The existing loop is untouched, so sessions actually using the wheel take the same path as before.

Note: This bounds rather than eliminates the catch-up. Since each slot stores its own deadline tx_tsc_, skipped slots would go stale. We still need to update tx_tsc_ once for each slot.

2. Fresh timestamp for progress_tsc_ (src/rpc_impl/rpc_req.cc)

enqueue_request() now stamps progress_tsc_ with dpath_rdtsc() instead of ev_loop_tsc_, so a newly enqueued request is never seen as already timed out by the packet-loss scan. Costs one rdtsc (~8 ns) per enqueue_request.

Evaluation

Two-node CloudLab xl170 cluster, RoCE 25 Gbps, 16-byte request/response, one outstanding RPC. Client enqueues, polls to completion, sleeps for wait. Mean end-to-end latency; retransmissions went from one per RPC to zero at every wait above the RTO.

wait	before	after
0	4.68 µs	4.70 µs	—
10 ms	386 µs	41 µs	9×
100 ms	3,690 µs	38 µs	97×
1 s	36,625 µs	56 µs	654×
5 s	75,790 µs	60 µs	1,270×

Latency is now flat in the idle period rather than linear. reap() in isolation (no NIC, empty wheel) drops from 149 ms to 24 µs after a 10 s idle, and is flat ~10–25 µs across the range. No change on the zero-wait path that §6.2's overhead numbers measure.

Testing

All relevant unit tests pass (PERF=OFF): timing_wheel_test (including TimingWheelRateTest — pacing still achieves its 1/5/10/20/40 Gbps targets), packet_loss_test, rpc_kick/req/cr/rfr/resp_test, small_msg_test, large_msg_test.