[wu-sheng]

Fix untraced Lettuce batch writes, orphan reactive traces, and unfinished spans of done commands

Follow-up to #841. Fixes three Lettuce plugin problems found while reviewing it. The goal: every Redis
server access gets one exit span. Linking it to a parent stays best effort and depends on upstream plugins.

Bug fixes

1. Batch writes were never traced.

• Cause: RedisChannelWriterInstrumentationV5/V65 matched write(List). Lettuce declares
  write(Collection<? extends RedisCommand>) in every RedisChannelWriter, so commands sent with
  StatefulConnection#dispatch(Collection) (also used by Redis command interfaces with @BatchSize) produced no span.
• Fix: match write(Collection). The existing interceptor then builds one Lettuce/BATCH_WRITE exit span for the
  whole flush.
  • The span is carried by the last command that is not done, since Lettuce drops done commands but still sends the rest.
  • Nested writers (expiry, listener, master/replica) reuse that span.
  • A cluster batch gets one span per node write, the same way Lettuce splits it.
• Scope: the span records only cache.type and cache.cmd=BATCH_WRITE. Like the Jedis pipeline span, it has no
  per-command details.

2. Every reactive command created without an active trace started a trace of its own.

• Cause: in WebFlux, the createMono/createFlux interceptor runs on a thread with no active context, because the
  webflux plugin hands its entry span over to the Reactor Context. So ContextManager.createLocalSpan started a new
  trace: one orphan Lettuce/Reactive/createMono root segment per command. The snapshot it captured was then ignored,
  because webflux's snapshot was already in the Reactor Context.
• Fix: the interceptor does nothing when no context is active. A snapshot that another plugin put into the Reactor
  Context is still continued when the command is written. A command with no parent becomes a root trace holding only
  its exit span.

3. A command that was already done when written got a span that never finished.

• Cause: the CommandHandler drops a done command without another terminal call. A command can be done by then if
  it was cancelled or timed out while its write was delayed, for example by MasterReplicaChannelWriter waiting for a
  connection. A span created for it never finished, so its segment was never reported.
• Fix: the writer interceptor skips a command when either the written command or its span carrier is done.
  • It re-checks after setting the span, for a cancellation that ran while the span was created.
  • Whoever finishes the span now takes it atomically (RedisCommandEnhanceInfo#takeSpan), so concurrent
    complete/cancel/re-check finish it exactly once. Before, they could finish it twice.

Known limitations

• A batch member cancelled while the batch is being written can leave that batch without a span, or with two finished
  spans. Closing this needs one span shared by all batch members, which is left for a later change.
• Lettuce 5.0/5.1 only: Command#completeExceptionally does not set the command's status in those versions. A reactive
  command that times out while a master/slave write waits for a connection can therefore still leave a span unfinished.
• Not changed, by decision: the connection handshake (RedisHandshake writes HELLO/AUTH/SELECT/CLIENT straight to the
  channel) stays untraced. Manual pipelining (setAutoFlushCommands(false) + flushCommands()) keeps one span per
  command.

Tests

• lettuce-scenario and lettuce-6.5.x-scenario dispatch a two-command batch and expect one Lettuce/BATCH_WRITE span.

• lettuce-webflux-5x-scenario and lettuce-webflux-6x-scenario assert the exact segment count: 5 (health check, entry,
  three Redis commands). Before this fix there were 8.

• RedisChannelWriterInterceptorTest adds unit tests for the cases a scenario cannot reproduce: a done command,
  cancellation while the span is created, two terminal interceptors racing for one span, and a done last command in a
  batch.

• Ran locally for every version in the four support-version.list files (16 versions): mock collector, agent,
  redis:3.2.9-alpine, validator. All pass with this change. All fail with main's Lettuce plugin jars: no
  BATCH_WRITE span, and 8 webflux segments instead of 5. The new unit tests also fail against main's code.

• Add a unit test to verify that the fix works.

• Explain briefly why the bug exists and how to fix it.

• Update the CHANGES log.