Update user-facing documentation to reflect PR #2 implementation.

Changes:
- Add comprehensive "Sampling for Production" section with 3 strategies
- Document new --journey-tracing-sample-rate flag (default 1.0)
- Explain vLLM native sampling vs OTEL sampling vs collector sampling
- Add comparison table for choosing the right sampling strategy
- Update configuration examples with sampling use cases
- Add Technical Details section on sampling architecture
- Add FAQ entries: vLLM vs OTEL sampling, atomicity guarantees
- Update Performance Impact section with sampling overhead details
- Update troubleshooting section with vLLM sampling solutions
- Add early mention of sampling capability in introduction

Key messages for users:
- Default behavior unchanged (sample_rate=1.0, backward compatible)
- vLLM native sampling reduces all overhead (recommended for production)
- End-to-end atomic: either both spans exist or neither (no partial traces)
- Independent from OTEL traceparent sampled bit
- Recommended rates: 10% for 1K-10K RPS, 1% for >10K RPS

Critical fixes:
- Fix service name vs tracer scope confusion in Jaeger navigation
  (service.name is what users select, scope.name is span attribute)
- Correct AsyncLLM span creation claims (was: "creates only core span",
  now: "creates no spans by default, core-only if manual header set")
- Eliminate contradiction: early doc claimed AsyncLLM creates spans,
  later sections correctly said no spans without manual header
- Qualify "every request creates two spans" to "when using vllm serve"
- Qualify sampling sections to explicitly state vllm serve requirement

Accuracy improvements:
- Soften overhead numbers: "~200-300ns" → "sub-microsecond" (less brittle)
- Qualify authority model as "OpenAI API Server" (not generic "API layer")
- Add comprehensive AsyncLLM FAQ with working code examples
- Add deployment modes section distinguishing vllm serve vs AsyncLLM

Impact: Prevents user confusion about AsyncLLM behavior (expecting
automatic tracing → getting zero traces → filing bugs). Documentation
now accurately reflects codebase reality verified in scheduler.py and
test_journey_tracing_sampling.py.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Non-streaming completion requests (/v1/completions with stream=false) were
missing all _finalize_api_span() calls, causing llm_request spans to never
export to OTLP collectors. This resulted in incomplete traces with only
llm_core (engine layer) spans visible, while llm_request (API layer) spans
remained orphaned in memory.

Root cause: The non-streaming code path (lines 319-368) had no finalization
on success, error paths, or fake stream generator (beam search with stream=true).

Added comprehensive span finalization matching the pattern used in streaming
completions and chat completions:
- Error paths: Finalize with ABORTED for CancelledError, GenerationError, ValueError
- Fake stream generator: Added try-finally with DEPARTED before [DONE]
- Success path: Finalize with DEPARTED before returning response
- Outer finally block: Unconditional cleanup for any uncaught exceptions

Impact:
- Fixes: Non-streaming /v1/completions now exports complete API-layer traces
- Preserves: Streaming completions continue to work (no changes to that path)
- Matches: Behavior now consistent with /v1/chat/completions endpoint

Testing:
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen/Qwen2.5-0.5B", "prompt": "Test", "max_tokens": 20}'

Expected result: Both llm_request (scope: vllm.api) and llm_core
(scope: vllm.scheduler) spans now appear in OTLP traces with proper
parent-child relationship.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Adds ts_monotonic_ns field to RequestJourneyEvent for improved timestamp
precision. Uses single clock read with exact consistency (derive float from
int) to ensure both ts_monotonic and ts_monotonic_ns represent identical
instant. Fully backward compatible with default value of 0 for legacy code.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Step tracing work is complete. Removing planning document.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Removes all float equality comparisons (e.g., assert ts.monotonic == value)
from integration tests. Tests now only verify:
- Presence of both timestamp fields
- Type correctness (float/int)
- Exact consistency via integer round-trip validation

This ensures robustness against float precision issues as specified in
the PR #1 constraints.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>