fix(streaming): use monotonic clamped latency timings

[sayonfortify]

Standardizes FortifyRoot streaming latency elapsed-time measurement for TTFT/STTG across the SDK provider wrappers.

What changed:

• OpenAI: use time.perf_counter() for chat wrapper streaming/duration elapsed math; clamp elapsed seconds/ms at zero.
• Anthropic: use perf_counter for streaming start/first/end clocks; clamp TTFT/STTG and duration histogram elapsed values.
• Bedrock: use perf_counter for stream start/first/end clocks in invoke/converse streaming safety wrappers; clamp TTFT/STTG values.
• LiteLLM: use perf_counter and clamp elapsed seconds before setting wrapper + canonical attempt span TTFT/STTG attrs.
• LangChain: SpanHolder start_time uses perf_counter; callback TTFT/STTG and duration math clamps non-negative.
• Google GenAI and LlamaIndex already used perf_counter + clamp; untouched.

Why:

• TTFT/STTG are durations and should not use wall-clock time that can jump under NTP/system clock changes.
• Span attrs should never carry negative latency values.

Tests:

• python3 -m py_compile on all changed source + test files.
• conda run -n py3.12 python -m pytest packages/opentelemetry-instrumentation-openai/tests/test_streaming_latency_span_attrs.py -q -> 4 passed
• conda run -n py3.12 python -m pytest packages/opentelemetry-instrumentation-anthropic/tests/test_safety_unit.py -k 'streaming_latency or latency_span_attrs or clamps_negative_duration' -q -> 4 passed
• conda run -n py3.12 python -m pytest packages/opentelemetry-instrumentation-bedrock/tests/test_safety_hooks.py -k 'streaming_latency or latency_span_attrs or clamps_negative_duration' -q -> 3 passed
• conda run -n py3.12 python -m pytest packages/opentelemetry-instrumentation-litellm/tests/test_safety_hooks.py -k 'streaming_latency or latency_attrs or elapsed_seconds_clamps' -q -> 4 passed
• conda run -n py3.12 python -m pytest packages/opentelemetry-instrumentation-langchain/tests/test_streaming_latency_span_attrs.py -q -> 3 passed