Use SWRX timestamps in server tests to match production

[leoleovich]

Summary:
TestServer fails non-deterministically (and consistently on newer kernels,
e.g. Fedora rawhide where it breaks the RPM %check stage) with a UDP read
timeout at server_test.go:223.

Root cause: the tests configure the server with TimestampType: timestamp.SW,
a value production never uses. cmd/ntpresponder defaults to SWRX and
Config.Validate() only accepts SWRX/HWRX. SW routes through
EnableSWTimestamps, which additionally turns on TX software timestamping
(SOF_TIMESTAMPING_TX_SOFTWARE). Every response the worker sends then makes
the kernel clone the packet onto the socket error queue. The responder never
drains MSG_ERRQUEUE, so those clones accumulate; error-queue skbs are charged
to sk_rmem_alloc and bounded by SO_RCVBUF. Once that budget fills, the kernel
silently drops incoming requests, the listener never sees them, and the
client's synchronous read blocks until the 10s deadline. How quickly it fills
relative to the 1000-request loop depends on skb truesize / rmem accounting,
which is why it passes on some kernels and fails on others; the earlier 1MB
buffer bump only delays the fill.

Measured difference (1MB rcvbuf, 4000 sends, no draining):

Mode	sk_rmem_alloc after sends
SWRX (production, RX only)	0 bytes
SW (test, TX+RX)	2,096,640 bytes (entire 2MB rcvbuf)

Switch TestServer and TestListener to timestamp.SWRX so the tests exercise
the real production code path and no longer accumulate TX timestamps.

Differential Revision: D107533486

[meta-codesync]

@leoleovich has exported this pull request. If you are a Meta employee, you can view the originating Diff in D107533486.

[meta-codesync]

This pull request has been merged in 67daa0f.