[AGENT-16861] Fix silent log loss on transient network errors

[ddrthall]

Summary

fluent-plugin-datadog silently dropped in-flight log batches on transient network interruptions. Two gaps combined:

1. On the HTTP transport, raw network exceptions (Net::ReadTimeout, EOFError, Errno::ECONNRESET, etc.) raised by the persistent client were not classified as RetryableError, so the plugin's own send_retries never retried them. (Ruby's Net::HTTP won't retry a POST for us — POST is non-idempotent.)
2. write(chunk) wrapped its body in rescue Exception => e and only logged, returning normally. Because a normal return signals success, Fluentd core committed and purged the chunk — its buffer retry never engaged.

Net effect: on a transient blip, a batch of records was dropped with no retry from either the plugin or Fluentd core, and no operator-visible failure.

Changes

• Classify transient network errors as retryable in DatadogHTTPClient#send (mirrors the exception set Net::HTTP retries for idempotent requests, plus Net::HTTP::Persistent::Error), routing them through the existing send_retries backoff loop.
• Re-raise from send_retries once bounded retries are exhausted, instead of returning normally and silently dropping the payload.
• Log then re-raise in write (narrowed from Exception to StandardError, so SystemExit/Interrupt are not trapped) so Fluentd core's buffer retry/backoff engages as designed.

The two retry layers are complementary: plugin-level send_retries retries only the failing batch (avoiding duplicate re-sends of already-accepted batches within a multi-batch chunk), while Fluentd core provides durability, backpressure, and bounded retry lifetime as the backstop.

Test plan

• bundle exec rake test — 44 tests, 0 failures
• HTTP send raises RetryableError on Net::ReadTimeout, EOFError, Errno::ECONNRESET
• send_retries re-raises after bounded exhaustion; does not retry non-retryable errors
• write propagates rather than swallows a client failure

Fixes AGENT-16861.