Fix ingest compatibility metadata gaps

[haasonsaas]

Summary

This fixes three ingest compatibility gaps I found while reviewing the core SDK/OpenLLMetry path:

• Preserve legacy Cohere/OpenLLMetry prompt content emitted as gen_ai.prompt.N.user when normalizing indexed gen_ai.prompt.* messages.
• Honor top-level raindrop.replayRunId, which parse.ts documented as supported but did not actually read.
• Persist and display total-only token usage from attributes such as llm.usage.total_tokens, instead of dropping it when split input/output token counts are unavailable.

I searched existing issues/PRs for gen_ai.prompt.0.user, Cohere, replayRunId, and total_tokens. I did not find an existing issue/PR for these follow-ups. The only related open PR is #19, which covers the separate legacy gen_ai.usage.prompt_tokens / gen_ai.usage.completion_tokens mismatch.

Details

• Adds spans.total_tokens plus the generated Drizzle migration/assets.
• Carries total_tokens through ingest, cloud/saved run mapping, API schemas, span skeletons, filtering/sorting, and run/detail displays.
• Keeps split input/output tokens as the source for existing cost calculation; total-only counts are shown as token metadata but not treated as split cost inputs.
• Adds parser tests for top-level replay IDs, Cohere .user prompt normalization, and total-only usage attributes.

Validation

• bun test tests/parse.test.ts
• bun test tests/
• bun x tsc --noEmit
• bun run build
• bun run lint (passes with existing warnings)