Update failure-finish-reason-safety with Vertex responses

[andrewheard]

Replaced streaming-failure-finish-reason-safety.txt, unary-failure-finish-reason-safety-no-content.json and unary-failure-finish-reason-safety.json with responses from Vertex AI. The previous ones appear to be sourced from Google AI (example). These responses have a few new fields in safetyRatings:

• probabilityScore
• severity
• severityScore
• blocked

Additionally the modelVersion is included in the response, and index is omitted. These responses also did not include promptFeedback.

[github-actions]

Coverage Diff

GenerateContentResponse: 43
| candidates: 38
| | index: 35 -> 32 🟡
| | safetyRatings: 29
| | | blocked: 0 -> 3 ✅
| | | probability: 29
| | | | LOW: 1 -> 4 🔵
| | | | HIGH: 4 -> 1 🟡
| | | probabilityScore: 7 -> 10 🔵
| | | severity: 7 -> 10 🔵
| | | | HARM_SEVERITY_NEGLIGIBLE: 7 -> 10 🔵
| | | | HARM_SEVERITY_LOW: 0 -> 3 ✅
| | | severityScore: 7 -> 10 🔵
| modelVersion: 0 -> 3 ✅
| promptFeedback: 27 -> 24 🟡
| | safetyRatings: 26 -> 23 🟡
| | | category: 26 -> 23 🟡
| | | | HARM_CATEGORY_HATE_SPEECH: 24 -> 21 🟡
| | | | HARM_CATEGORY_DANGEROUS_CONTENT: 25 -> 22 🟡
| | | | HARM_CATEGORY_HARASSMENT: 26 -> 23 🟡
| | | | HARM_CATEGORY_SEXUALLY_EXPLICIT: 24 -> 21 🟡
| | | probability: 26 -> 23 🟡
| | | | NEGLIGIBLE: 25 -> 22 🟡
| usageMetadata: 12 -> 15 🔵
| | candidatesTokenCount: 11 -> 12 🔵
| | promptTokenCount: 12 -> 15 🔵
| | totalTokenCount: 11 -> 14 🔵
Total Coverage: 39.29% -> 41.43% ✅

Legend:
✅ : total coverage increase
🔵 : files number increase, total coverage unaffected
🟡 : files number decrease, total coverage unaffected
❌ : total coverage decrease

[andrewheard]

@hsubox76 @DellaBitta The previous sample files came from Google AI where the index field was always included. Vertex AI omits default values (index == 0) when serializing its responses as JSON. I filed firebase/firebase-js-sdk#8566 but @dlarocque can likely explain it much better than me (thank you for the debugging help, Daniel!!).