fix: ensure prompt_tokens is always present in Anthropic streaming responses

[vrushankportkey]

Summary

Fixes an issue where prompt_tokens was missing from streaming responses for Anthropic models (both direct Anthropic API and via Vertex AI), causing billing/usage tracking to be incorrect.

Root cause: When streaming, the message_start event may not always include input_tokens. The code relied solely on this event to set prompt_tokens, resulting in undefined values that were omitted from the JSON response.

The fix:

• Checks for input_tokens in the message_delta event as a fallback (per Anthropic docs, these are cumulative counts)
• Defaults to 0 if neither source provides the value
• Explicitly sets prompt_tokens in the usage object to ensure it's always present

Changes

• src/providers/google-vertex-ai/chatComplete.ts - Vertex AI Claude streaming
• src/providers/anthropic/chatComplete.ts - Direct Anthropic streaming

Test plan

• Test streaming responses from Anthropic /v1/chat/completions - verify prompt_tokens is present
• Test streaming responses from Vertex AI Claude /v1/chat/completions - verify prompt_tokens is present
• Test with tool use / structured output calls specifically (the reported issue scenario)
• Verify total_tokens calculation is still correct

Before (problematic response)

"usage": {
  "completion_tokens": 52,
  "prompt_tokens_details": {
    "cached_tokens": 0
  }
}

After (fixed response)

"usage": {
  "prompt_tokens": <value>,
  "completion_tokens": 52,
  "total_tokens": <calculated>,
  "prompt_tokens_details": {
    "cached_tokens": 0
  }
}