Summary

The speculative decoding pipeline computes acceptance stats and logs them to
console via pretty_format(), but never publishes them to the telemetry
system. This PR bridges that gap by adding three OTEL counters at DETAILED
metric level:

• maxserve.speculative.draft_tokens_generated
• maxserve.speculative.draft_tokens_accepted
• maxserve.speculative.bonus_tokens_used — this field already existed on
  SpeculativeDecodingMetrics but was never propagated to BatchMetrics.
  Adding it required touching the dataclass, create(), and the test
  constructor. It could be dropped deriving it from
  the other two counters is preferred.

BatchMetrics.publish_metrics() emits these counters only when
draft_tokens_generated > 0, avoiding noise when speculative decoding
is not active.

Additionally, SpeculativeDecodingMetrics is now reset after each batch in
BatchMetrics.create(), matching the existing kv_cache.reset_metrics()
pattern. Without this reset, BatchMetrics.create() reads cumulative
lifetime totals rather than per-batch deltas. This was already an issue
for pretty_format() console logs (which showed ever-growing totals
instead of per-batch stats), and would cause the new OTEL counters to
double-count since counters are additive.

Note: This changes the existing pretty_format() console log output
from cumulative lifetime stats to per-batch stats (e.g., "Draft Tokens:
15/30" instead of an ever-growing total). Every other metric in the log
line (batch size, throughput, KV cache stats) was already per-batch —
the speculative decoding fields were the only cumulative outlier.

Testing

• test_create_resets_speculative_metrics_between_batches — regression
  test that calls BatchMetrics.create() twice with the same
  SpeculativeDecodingMetrics object. Verifies the second call returns
  only the new batch's values, not cumulative totals. Fails without the
  reset() fix (batch 2 reads 30 instead of 10).

Checklist

• PR is small and focused
• I ran ./bazelw run format to format my changes
• I added or updated tests to cover my changes
• If AI tools assisted with this contribution, I have included an
  Assisted-by: trailer in my commit message or this PR description

Assisted-by: AI