fix(gemini,groq): include finish reason in unhandled finish reason exceptions

[datashaman]

Description

Enriches the default PrismException in Gemini\Handlers\Text, Gemini\Handlers\Structured, and Groq\Handlers\Text to include both the mapped FinishReason enum value and the raw provider finish reason string — matching the approach already used in OpenAI handlers (PR #941).

Previously threw a generic "Gemini: unhandled finish reason" with no indication of what the actual reason was (e.g. SAFETY, LANGUAGE, etc.). Now throws 'Gemini: unhandled finish reason "content-filter" (raw: SAFETY)'.

Relates to #205.

Changes:

• src/Providers/Gemini/Handlers/Text.php — enriched default exception with sprintf
• src/Providers/Gemini/Handlers/Structured.php — same
• src/Providers/Groq/Handlers/Text.php — same
• Added Pest tests for Gemini Text and Structured content filter exceptions
• Added test fixtures simulating SAFETY finish reason responses

Breaking Changes

None — only changes exception message strings in error paths.

[datashaman]

Additional context: unhandled finish reasons

While investigating this fix, I catalogued the complete set of finish reasons each provider can return vs what's currently handled.

Gemini

Gemini's FinishReasonMap currently maps 10 values, but Google's API defines at least 20+. These newer values are not in the map (they fall through to default => FinishReason::Other, then hit the exception this PR improves):

Value	Notes
IMAGE_SAFETY	Image generation safety violation
IMAGE_PROHIBITED_CONTENT	Image generation prohibited content
IMAGE_RECITATION	Image generation recitation
IMAGE_OTHER	Image generation misc
NO_IMAGE	Expected image but none produced
UNEXPECTED_TOOL_CALL	Tool call generated but no tools enabled
TOO_MANY_TOOL_CALLS	Too many consecutive tool calls (Google AI only)
MISSING_THOUGHT_SIGNATURE	Thought signature missing (Google AI only)
MALFORMED_RESPONSE	Malformed response (Google AI only)
MODEL_ARMOR	Blocked by Model Armor (Vertex AI only)

Additionally, 8 values that are in the FinishReasonMap (mapping to ContentFilter or Other) still hit the default exception branch in the handler because only Stop, Length, and ToolCalls are matched:

SAFETY, RECITATION, BLOCKLIST, PROHIBITED_CONTENT, SPII, MALFORMED_FUNCTION_CALL, LANGUAGE, OTHER, FINISH_REASON_UNSPECIFIED

Groq

Groq only has 4 finish reasons (stop, length, tool_calls, function_call). The first three are handled. function_call (deprecated) is not in the FinishReasonMap.

---

These are potential follow-up items — this PR just ensures the exception messages are useful for debugging when any of them are encountered.

Sources: Google AI REST v1, Vertex AI REST v1, Groq API reference