feat: Add completion start time timestamp to relevant generators

[vblagoje]

Why:

Enhances Langfuse integration to track the time-to-first-token (TTFT) in streaming responses. The new entry in chat message metadata completion_start_time enables TTFT tracking in Langfuse. We'll separately update all other relevant generators in https://github.com/deepset-ai/haystack-core-integrations/

• fixes Langfuse integration: completionStartedAt is not logged, required for streaming metrics like TTFT and TPM/TPS haystack-core-integrations#1261

What:

• Added completion_start_time metadata in the Hugging Face API and OpenAI chat and non-chat generators to record the TTFT.
• Modified test cases to validate the inclusion of the completion_start_time and its format correctness.

How can it be used:

This change allows developers to track the exact start time of streaming completions, aiding in performance diagnostics

How did you test it:

Included updated unit tests for both Hugging Face API and OpenAI chat and non-chat generators to ensure the presence and correct ISO format of completion_start_time metadata. These tests verify that the timestamps are recorded and formatted correctly during streaming responses.

Notes for the reviewer:

Ensure close attention to the new metadata handling and its correctness across different components. Testing has validated the format, but real integration checks are recommended for full assurance.

[coveralls]

Pull Request Test Coverage Report for Build 12825406823

Details

• 0 of 0 changed or added relevant lines in 0 files are covered.
• 13 unchanged lines in 4 files lost coverage.
• Overall coverage increased (+0.009%) to 91.301%

---

Files with Coverage Reduction	New Missed Lines	%
components/generators/hugging_face_api.py	2	96.3%
components/generators/chat/hugging_face_api.py	3	96.46%
components/generators/chat/openai.py	4	96.3%
components/generators/openai.py	4	95.65%

Totals
Change from base Build 12794162077:	0.009%
Covered Lines:	8848
Relevant Lines:	9691

---

💛 - Coveralls

[vblagoje]

@julian-risch @anakin87 while at this, I suggest we do the same for the popular generators integrations like ollama, bedrock, and cohere. Or perhaps not - LMK.

[anakin87]

I'm not a fan of introducing metadata that are very specific (for Langfuse in this case), but I don't see alternative solutions and the PR looks good.

On introducing this metadata in other (Chat)Generators, I would only focus on those supported by our Langfuse implementation: https://github.com/deepset-ai/haystack-core-integrations/blob/72989f4461b863d8be6051dbed5fdefdd6d870ed/integrations/langfuse/src/haystack_integrations/tracing/langfuse/tracer.py#L18